Obsidian MCP Server

Obsidian MCP Server

[Python] [License] [MCP] [PostgreSQL]

Your agents' memory, in the vault you already have, reachable from claude.ai, ChatGPT, Claude Code and your phone.

A self-hosted Model Context Protocol server over an Obsidian vault. Three things it does that the alternatives mostly don't:

• Your existing vault, served in place. No imposed note format, no sidecar syntax, no "memory" folder you have to adopt. Agents read what you read, follow your wikilinks, and pick up your folder layout, frontmatter schema and tag conventions from the vault itself.
• Remote and multi-user. OAuth 2.0 PKCE and API keys, so claude.ai, Claude Desktop, Claude mobile, ChatGPT, Claude Code and n8n all read and write the same files, each user confined to their own vault root.
• Writes cannot clobber. expected_hash preconditions refuse an edit made from a stale read, every publish is atomic and no-clobber by default, deletes go to .trash/, and dry_run returns the diff before anything lands.

Connect from Claude Code:

claude mcp add obsidian --transport http \  --url "https://obsidian-mcp.<your-domain>/mcp" \  --header "Authorization: Bearer omcp_..."

In claude.ai, Claude Desktop or ChatGPT, add a custom connector at https://obsidian-mcp.<your-domain>/mcp and sign in through the OAuth consent page. Connect a client has the rest.

Stack: Python 3.12, FastAPI, PostgreSQL with pgvector. Pluggable embeddings (Ollama — bge-m3 by default, qwen3-embedding:0.6b in the reference deployment — or OpenAI text-embedding-3-{small,large}).

[Dashboard]

Contents

• Why this exists
• A session at the keyboard
• A session away from the keyboard
• What's in the box
• vs. other Obsidian MCP servers
  • vs. hosted memory systems
  • vs. an agent with raw file access
• Who this is for
• Control panel
• Quick start
• Versioning and compatibility
• Upgrading
• Cost expectations
• The self-describing vault
• Multi-user mode
• Configuration
• Architecture
• Project layout
• Development
• Security notes

Why this exists

There are three things going on here, and they're more interesting together than apart.

1. Agent memory that you can actually read

If you let an agent run for a while, it needs memory. Most setups solve this with an opaque vector store, a SQLite blob, or a managed "memory" service that you can't see into. That works until you want to know what the agent thinks it knows about you, or you need to correct something, or you want to understand why it just made a weird suggestion.

This server gives you a different deal. Agent memory lives as markdown files in your vault. Folder structure, file names, frontmatter, all visible. You can open the file in Obsidian and read it. You can edit it. You can delete it. You can grep it. The agent's "memory" is a human-auditable artifact that sits in the same place as your own notes, with the same tools available.

The home lab is the use case that sold me on this. My vault has notes on the rack, the network, and every Home Assistant integration. I can say "set up a night-light mode in the master bathroom, 1% after 11pm" and a sysadmin agent finds the right config, makes the change, and updates the doc in the same pass. Six months later when I've forgotten how it works, the answer is in the vault, not buried in some chat history I can't search.

The semantic search and wikilink graph still work over that material, so retrieval is fast and conceptual. But the substrate is files you own, not a black box.

To be precise about the scope: what the server supplies is MCP-accessible storage, keyword and semantic search, and graph operations over markdown notes, for whatever MCP clients you connect. The agents direct their own reads and writes. There is no extraction or decay pipeline running behind them, and nothing rewrites a note because a model judged it stale: an agent remembers something because it wrote a note, and forgets it because someone deleted one. By late 2026 that is less of a contrarian position than it sounds. Every vendor's consolidation step (the "dreaming" or "sleep-time" job) has turned into a separate process you can inspect, rather than silent extraction inside the write path. This server keeps the agent as the author. The consolidation help it will grow is a hygiene report that proposes merges and flags duplicates, and never writes (#359), not a pipeline that edits your notes while you are not looking.

2. A shared memory layer between you and your agents

The other half runs the other way: the vault isn't only the agents' memory, it's mine. I think of my Obsidian vault as my exocortex. The "big me" that includes notes, calendars, scripts, search, and AI assistants is substantially more capable than the "small me" of the biological brain alone. It's also where I do most of my thinking, because writing something down is itself a form of thought.

The problem is that until recently, the vault was passive. I had to go find things. Agents that wanted to help me had to be briefed from scratch every session, and they had no way to see what I'd already written about a topic.

This server fixes that. Now the same vault feeds my own daily writing and any agent I plug into it. The agent reads what I read, links what I link, follows the same wikilinks, sees the same frontmatter. When I write a project note on Sunday, my Monday-morning briefing agent already knows about it. When the agent leaves notes from a research session, they show up in my normal Obsidian search.

A concrete version of this: I'll spend a session in Claude Code on a project, wrap up, push the commits, and then just say "update Obsidian." The agent reads the vault guide, figures out where project notes live in my structure, picks the right format and frontmatter, and leaves a session log I can later roll into a status report. No path-passing, no telling it what to write — the conventions are already in the vault, and it follows them.

That's the exocortex idea made concrete: one place that holds context, and both the human and the agents reading and writing into it on the same terms.

3. The vault follows you

The thing that still surprises me is that this is internet-facing. Same vault, same notes, reachable from anywhere I happen to be talking to an agent — Claude in a browser tab, Claude on my phone on the train, an n8n workflow firing on a schedule, a Claude Code session on whatever laptop is in front of me. All behind OAuth or a bearer key, all hitting the same source of truth.

So when I hear an interesting podcast on a commute and want to think out loud about it with Claude on my phone, the agent isn't starting cold. It can pull up what I've already written on adjacent topics, surface a note I half-finished six months ago, and at the end of the conversation suggest updates and write them in. The vault doesn't have to be near me to be the thing I'm working in. (A session away from the keyboard walks through exactly one of these.)

A session at the keyboard

To make the abstract concrete, a short transcript from a real session. Tool calls inline; the user types in plain English and the agent picks the right tool.

you  > find anything I've written about why retrieval-augmented       chat is hard
claude  → semantic_search(query="why retrieval-augmented chat is hard")    • 2025-12-03 - RAG failure modes.md    • 2025-10-18 - Chunking tradeoffs.md    • 2024-06-22 - Vector recall vs precision.md
  the closest hit is your December note on RAG failure modes. let  me see what it's connected to.
  → get_neighborhood(path="2025-12-03 - RAG failure modes.md",                     depth=1)    8 connected notes. notable ones:    • "Hybrid search" — a re-ranking design you sketched but      didn't ship    • "Out-of-distribution queries" — a half-finished draft from      October    • "Embedding model choice" — links to bge-m3, your current      model
  you have three connected pieces here: a failure-modes catalog,  an unshipped hybrid-search sketch, and an abandoned O-O-D draft.  want me to summarize all three together, or pull up the draft  so you can decide whether to finish it?
you  > pull up the draft, and tag it #revisit-2026 in frontmatter       while you're there
claude  → read_note(path="2025-10-18 - OOD queries (draft).md")  → set_frontmatter(      path="2025-10-18 - OOD queries (draft).md",      updates={"tags": ["draft", "rag", "revisit-2026"]}    )
  here's the draft, and the tag is on. you stopped right before  the section on confidence thresholds; the open question you  left yourself was…

Two things to notice. First, the agent didn't need to be told what folder to look in or what tools to use — it picked them. Second, the write at the end is structured (set_frontmatter mutating YAML, not a regex over the file body), so the note round-trips cleanly. The self-describing vault and the wikilink graph are doing the work that makes this feel natural.

A session away from the keyboard

The transcript above is the easy case: I'm at a desk, I can see what the agent is doing, and Obsidian is one alt-tab away. The session that actually changed how I think about this server had none of that.

I was out walking with a health podcast in my ears — a long one, two people who clearly disagreed with each other, an hour of it. I had my phone and no intention of going home to a laptop. So I pulled the episode's transcript, handed it to Claude on my phone, and we talked it through while I kept walking: what the actual claim was, which parts I already had notes on, where it cut against something I'd decided months ago and written down at the time.

The agent had the vault the whole way. It surfaced what I'd already written on the topic, flagged that two dates in an older note were wrong, and asked whether a decision I'd recorded last year still stood given what the episode argued. By the time I got back it had written all of it in: the health-related decisions I'd actually landed on during the walk, the date corrections in the old note, a couple of new notes on the episode itself — and, because the conversation kept circling back to it, a durable note on how I decide which experts to trust on medical questions in the first place. That last one is the artifact I keep returning to. It wasn't about the episode at all; it was the reasoning underneath a whole class of decisions, and it now sits in the vault where the next agent will find it.

I never opened Obsidian. Not on the walk, not when I got home. The whole session — retrieval, argument, correction, and the writing that came out of it — went through an agent, and the vault is simply where it landed. Obsidian is how I check the work afterwards, not how the work gets done. That inversion is most of the reason this project looks the way it does.

What's in the box

The server exposes 28 MCP tools across five families, plus the auth and ops layer around them.

Every tool is published with a title and MCP ToolAnnotations, all four hints set explicitly: readOnlyHint on the eighteen tools that only read (search, read, list, graph, list_tasks, get_changes, check_upload), destructiveHint on the eight that can replace or remove something (edit_note, move_note, delete_note, set_frontmatter, write_file, delete_file, import_from_url, and request_upload, whose link can overwrite on redemption), openWorldHint on import_from_url only. create_note and request_download are additive. Clients that honour annotations run read-only tools without a per-call confirmation (Claude Code also runs them in parallel) and keep prompting on destructive ones. They are hints: the server enforces read-only credentials itself, regardless of what a client does with them. The initialize result also carries a short instructions paragraph — what the vault is, which tool to reach for, and the write-safety rules — for clients that load only tool names at session start. No description exceeds 4,000 characters on the wire, so none is cut off by a client's description limit.

Search and discovery

• keyword_search(query, folder?, tags?, frontmatter?, limit=20), full-text via PostgreSQL tsvector. A note's title and frontmatter aliases rank above its body, so a query that is a note's title finds that note first; its tags (frontmatter and inline #tags, a nested #a/b also as a and b) are indexed below the title and above the body, and matched_in names which of title, tags and body matched; bare terms are ANDed, and when the full matches do not fill the page, partial matches follow tagged match: any; every row carries a snippet: with the matches highlighted (or head:, the note's opening words, for a title or tags match). The text-search config(s) are configurable via FTS_CONFIGS (see Full-text search language(s))
• semantic_search(query, folder?, tags?, frontmatter?, limit=15, rerank=false), vector similarity via pgvector, one one-line preview chunk per note. Notes are chunked at their headings, and each chunk is embedded under a Title > Heading > Subheading context prefix. A fresh row names the section it matched: § "Plan/Risks" (#7) or § (top of note), and read_note(path, section="#7") reads exactly that section. Previews start on a word; a section too short to stand alone (under 200 characters) is grouped with its neighbour, so a preview can include an inner heading line and its § names the enclosing section. Stale rows, and rows not yet re-embedded since the upgrade to heading-aware chunks, carry no locator. While any of your notes is still on the old chunking, each response says so on an index_transition: line under its header
• hybrid_search(query, limit=10, folder?, tags?, frontmatter?, rerank=false), both of the above fused by reciprocal rank fusion over each arm's top 50, one line per note with its fused score, each arm's rank and the 1/(k+rank) terms behind the score. The three are peers, and every description carries the same routing sentence: exact identifiers, code symbols, titles or known phrases go to keyword_search; conceptual or paraphrased questions to semantic_search; unsure which, hybrid_search. A single arm can beat the fusion on its own kind of query. If the embedding provider is down, hybrid_search returns the keyword ranking marked arms: keyword-only (semantic unavailable: …), never silently degraded: the marker is on every return path, so a keyword-only page cannot pass for a fused one.
• Every page of semantic_search and hybrid_search carries a confidence: ok | low header token. low means no keyword note matched all the terms and the closest semantic candidate is under SEARCH_LOW_CONFIDENCE_SIMILARITY, with a one-line hint to try keyword_search or get_neighborhood when it costs no row. Nothing is filtered. The threshold ships at 0 (off: always ok) until you calibrate it.
• rerank=true on either of the last two asks an optional, operator-configured reranker (RERANK_URL) to reorder the candidates the call already fetched (at most 50 notes). Each note is scored on the best of up to four of its fetched chunks, sent as the same text they were embedded from (title, heading path and tags, then the chunk); a reranked row's preview and § locator show the chunk that scored. The response is either fully reranked (reranked: true (candidates: N), a rerank_score per row) or the unreranked page marked reranked: false (<reason>); without a reranker it answers reranked: false (not_configured). Its rows are one-line quoted objects, like hybrid_search's
• list_notes(folder?, limit=50, since?, until?, order_by?, cursor?), sorted by modified time
• get_recent(folder?, limit=20, since?, until?, order_by?, cursor?), recently changed, with each modification time in UTC (YYYY-MM-DD HH:MM UTC). On both, order_by="content_changed" with since lists notes whose content changed (a touch or a re-sync that keeps the bytes does not count); since is inclusive and until exclusive, each a UTC date (YYYY-MM-DD, the whole day) or an ISO 8601 datetime with an explicit offset. Any of the four switches to keyset paging with a more: cursor=… footer; a call with none of them is unchanged
• get_changes(since, folder?, kinds?, by="any", limit=50), the change feed: one line per note created, modified, moved or deleted after a cursor, oldest first, each attributed to the API key or OAuth client that made it ((you) for your own credential) or human/sync for an edit the index discovered (Obsidian, a sync client, git). It returns next_since, an exact ev:<n> cursor an agent stores in its own notes for the next session; by is any | me | agents | human. Deletions say why (removed, not_indexable, quarantined); gap and reset lines say when the view is incomplete
• get_tags(limit=50), tag and count
• list_tasks(folder?, status="open", due_by?, scheduled_by?, tags?, limit=100, cursor?), checkbox tasks (- [ ], - [x], - [-], - [/], …) across the vault from the index, never reading a note: one line per task with its path and line, Tasks-plugin dates (📅 ⏳ 🛫 ✅), priority, the task's own #tags, its heading path and its text. Ordered by due date (undated last), paged with a next cursor; due_by / scheduled_by mean "on or before". Tasks-plugin query blocks are not evaluated. The header always counts notes not yet scanned for tasks and notes whose tasks could not be indexed, so an empty list is never a silent "nothing open"
• get_vault_guide(), the Obsidian primer plus this vault's CLAUDE.md, served live
• Retrieval changes are measured with make search-eval against judgments mined from the operator's own usage log (a search followed by a read of one of its results); see the search evaluation harness

Read and write

• read_note(path, section?, offset=0, limit?) returns a structured result — path, title, tags, frontmatter_yaml and a JSON frontmatter view, heading (section reads), content, and truncation as data (truncated, offset, next_offset, total_chars, outline, notice). Bounded by MAX_READ_RESPONSE_CHARS (default 40,000) — see Response size limits. section=<heading> returns one section's body instead of the whole note; offset continues a truncated read.
• create_note(path, content), atomic write, refuses overwrite
• edit_note(path, …) with four mutually exclusive modes: full replace (default), append=True, find=… (with optional replace_all), or section=<heading> (ATX headings, supports Parent/Child path-style and #N ordinal disambiguation). dry_run=True returns a unified diff without writing. Legacy clients may use operation="append"; operation="replace" explicitly selects full replace.
• move_note(from_path, to_path, rewrite_links=False), relocates and optionally rewrites incoming [[Old]], [[Old|alias]], [[Old#anchor]], ![[Old]], and [[folder/Old]] references in source notes
• delete_note(path, permanent=False), soft-delete to .trash/<YYYYMMDD-HHMMSS>-<basename>-<8 hex> by default, via a single non-replacing rename, so it never overwrites an existing trash entry (a filesystem that cannot do that rename makes the soft delete refuse with a named error rather than fall back). permanent=True unlinks.
• set_frontmatter(path, updates, remove?), structured YAML mutation. Body is byte-identical when only frontmatter changes.

File access (non-markdown)

Raw read/write/browse of arbitrary vault files (PDFs, images, skill assets, data files) — distinct peers to the note tools, which stay markdown-only. Pure byte transport: no server-side PDF/text extraction, no embedding or indexing of non-markdown files.

• read_file(path, encoding="auto", offset=0, limit?), returns text-like files as text, images as an inline image block that renders in-client, and other binaries as a base64 string. text/base64 force the form. Refuses files over MAX_FILE_READ_BYTES (default 10 MB); text results are additionally bounded by MAX_READ_RESPONSE_CHARS and continue via offset. hash_only=True returns the whole file's content_hash without content; base64 results include that hash in their header. Text results remain plain text.
• write_file(path, content, encoding="base64", overwrite=False), lands a file in the vault; base64 for binary, text for UTF-8. No-clobber by default, auto-creates parent dirs, atomic write. Capped at MAX_FILE_WRITE_BYTES (default 25 MB).
• list_files(folder=".", pattern="*", recursive=False, limit=200), ls-style browse of files and subdirectories with size and mtime, glob-filterable and result-capped.
• delete_file(path, permanent=False), soft-deletes a non-markdown file to .trash/<YYYYMMDD-HHMMSS>-<basename>-<8 hex> with a single atomic rename. Refuses markdown (that is delete_note), directories, and symlinks.

All four reuse the path-traversal guard and exclude any path with a component starting with . (dot-directories and dot-files) (.obsidian, .git, .trash, …), matching the indexer's visibility rule.

Guarding edits against stale reads

Pass a read's content_hash as expected_hash when editing, updating frontmatter, moving or deleting a note, overwriting a raw file, or deleting a raw file. The canonical token is sha256:<64 lowercase hex>, computed over the complete raw file bytes; a section or truncated read_note still returns the whole-file hash. For raw files, use read_file(hash_only=True) or the base64 header. Do not hash the returned text yourself.

A stale token refuses the operation before mutation, with a final MCP-REFUSAL JSON line naming stale_precondition and the current hash. Re-read and reconsider the edit before retrying. Moves bind the source note only; moves and deletes still allow an in-place edit after their preflight comparison. Overwrites retain their separate in-call byte check. Successful publishing writes report a new hash when available.

The argument is optional by default. WRITE_PRECONDITION_REQUIRED=true requires it on the supported destructive calls; enable this only after clients supply tokens. Creation is exempt and refuses a supplied token as no_incumbent. Files above their read cap cannot be guarded.

File transfer

No MCP client can hand a tool the bytes of a file the user is looking at, so write_file is only usable when the agent already has the content. These tools close that gap with short-lived capability links, redeemed over the public /transfer/* routes.

• request_upload(path, overwrite=False, expires_in?), mints a single-use link bound to exactly one destination path. The human opens it, picks a file, and it lands at path — nothing else can be written with it.
• check_upload(upload_id), reports pending / uploading / completed (with path, size, sha256 and MIME) / unknown (a stream started and the server never recorded how it ended — read the path before re-minting) / revoked (the credential or vault root changed under the link) / expired, scoped to the identity that minted it.
• request_download(path, expires_in?), mints a link the human can save one vault file from. Usable more than once until it expires, and bound to the file's exact bytes at mint time.
• import_from_url(url, path, overwrite=False), fetches a public https asset straight into the vault under an explicit outbound deny policy (no private, loopback, link-local, metadata or tunnelled addresses, in any spelling, re-checked at every redirect).

The token travels in the URL fragment, which browsers never send, so no server-generated request target or access log contains it. Uploads are claimed before a body byte is read, published atomically with no-clobber semantics, and bound at mint time to the file state they were minted against — a link cannot silently undo an edit made while it was waiting. MCP_HOSTNAME or BASE_URL must be set; without a public origin the mint tools refuse rather than emit a localhost link.

Wikilink graph

• get_backlinks(path, limit=50), notes linking TO path
• get_links(path), outgoing links, both resolved and dangling
• get_neighborhood(path, depth=1, limit=50), undirected BFS over the resolved-link graph, capped at depth ≤ 5 and limit ≤ 200
• find_related(path, limit=10), semantic neighbors via averaged chunk embeddings and pgvector cosine distance, deduped per note, with the same section locator and index_transition: line as semantic_search
• find_orphans(folder?, limit=50), notes with zero in or out resolved links

Auth and ops

• API keys with the omcp_ prefix, stored as SHA-256 hashes, with read and readwrite permission scopes. Write tools refuse on read-only keys.
• OAuth 2.0 PKCE (S256) flow for public and confidential clients, including ChatGPT, Claude Desktop, and claude.ai. Dynamic registration defaults to both vault permission levels; the user chooses the actual grant on the consent screen. A read-only grant lists 19 tools — the nine write tools are left out of tools/list (and still refused if called by name); choose read + write at consent for all 28, and reconnect the client after changing a grant's scope.
• Control panel (Jinja2, hand-written CSS, vendored Chart.js, nonce-based CSP) for keys, usage logs, indexer status, embedding-provider info, and a danger-zone reset.
• Every tool call is logged to usage_logs with name, params (truncated to 200 chars), duration, response size, and the calling credential's name — recorded at call time, so the audit trail survives deleting the key or OAuth client it describes.
• /health is unauthenticated and returns status plus two capability fields: transfer_mount_check_available (the kernel supports the mount check transfer writes need) and vault_named_staging_fallback_active (a write has actually staged under a name on this process).
• /health also carries an indexer object (status, task_running, failing_scopes, embedding_failing_scopes, max_consecutive_failures, quarantined_notes, last_success_at), and the top-level status becomes "degraded" when the indexer task has died, any index, embedding or enumeration failure counter — or a scope's run of incomplete re-derives — reaches INDEXER_DEGRADED_AFTER_FAILURES, or a note is quarantined. The HTTP code stays 200 — a restart cannot repair vault content or a provider outage, and Kubernetes uses /health for liveness — so monitor the status field (a keyword or JSON check for "status":"ok"), not the status code. Counts only: no path, error text or user id.

Every write — note tools, write_file, uploads and imports — stages the new bytes in a temporary inode, fsyncs them, and only then publishes. Creation publishes with a kernel-atomic hard link that refuses to clobber; move_note and the soft delete publish with a single non-replacing rename; an overwrite is a same-directory rename onto the destination. The destination directory (and any directory the call created) is fsynced afterwards, so a crash mid-write can neither truncate a note nor lose one the server reported as written.

Staging happens in an unnamed inode wherever the filesystem supports one, so no temporary name is ever visible in the vault. On a mount that refuses that (some NFS exports do), those writes refuse with an error naming VAULT_ALLOW_NAMED_STAGING_FALLBACK; setting that flag takes named staging back on both write paths as a declared, weaker guarantee. See System requirements.

vs. other Obsidian MCP servers

The field has sorted itself into two shapes since this README was first written, and the axis that matters is not "lightweight versus heavy" but where the agent runs relative to the Obsidian app.

Desktop-local. The tool talks to a running Obsidian. The Local REST API plugin has shipped its own Streamable HTTP MCP server since 4.0.0 (2026-05-15), and 5.5.0 (2026-10-17) added If-Match conditional writes and result caps, so it is no longer "glue" but a real local MCP endpoint. The official Obsidian CLI (1.12) reads, writes, searches and walks backlinks, and requires the app to be running. In-app agent plugins sit in the same group. All of these get the app's own metadata cache for free; none of them work on a headless box, and none give claude.ai or a phone a way in without you exposing your desktop.

Headless and remote. The server works on the vault files directly, wherever they are mounted, with no Obsidian process. This server is one. Vault Cortex (aliasunder, September 2026) is a single container with OAuth 2.1, SQLite FTS5 plus sqlite-vec fused by RRF and a local reranker, tasks and daily-note helpers, and a bundled Obsidian Headless Sync client; one vault per container, no multi-user. StevenStavrakis/obsidian-mcp v2 (2026-08) is a stdio server with journaled atomic writes, etag/if_match preconditions and 30-day recovery snapshots, and no semantic search. jacksteamdev/obsidian-mcp-tools was archived on 2026-05-13 and is dropped from the table.

If Obsidian is open on the machine the agent runs on and you need no remote access, the official CLI or the REST plugin is the lighter choice. This server is for the headless, remote, multi-client and semantic case: nothing on the box but a container and Postgres, and every client you own hitting the same vault.

As of 2026-10-28:

The table reflects each project's documented features on that date, read from their READMEs and release notes; a ? means I could not confirm it either way. Verify the specifics before betting on them.

vs. hosted memory systems

The comparison that matters more, now that most of my vault traffic is agents rather than me, is against memory as a service: your agent calls an API, the service stores what it's told, and it hands back what it judges relevant later. mem0, Zep and Letta are the names people usually reach for. What follows is about that architecture — memory behind a service boundary — not about any one product's current feature list, which moves faster than a README can track.

One thing has changed since that paragraph was first true: files plus search is now the mainstream substrate, not the odd one out. Letta's MemFS is git-backed markdown; Claude Code's auto-memory is a MEMORY.md index plus topic files; Anthropic's memory tool is a file directory; OpenClaw and Hermes keep markdown workspaces. "Readable markdown memory" on its own no longer sets this server apart. The closest comparables are Basic Memory (markdown plus a SQLite or Postgres index, hybrid search, an optional reranker, structured frontmatter filters; its own folder and note format, AGPL, with a paid cloud) and doobidoo/mcp-memory-service (self-hosted, remote MCP over OAuth, opaque memory records rather than files). Against those, the differences are five, and plain:

1. The existing vault, served in place, with no imposed folder or note format.
2. Remote and multi-user over OAuth, one vault root per user.
3. Destructive-write guards: expected_hash, atomic no-clobber publish, .trash/, dry-run diffs.
4. MIT.
5. No hosted tier, and none planned.

The difference from memory-as-a-service is where the memory lives and who can open it.

• Readability. When memory sits behind a service API, reading it means whatever endpoint or console the service exposes, in whatever shape it stores. Here the memory is the artifact: Health/2026-08 - Trusting expertise.md, in a folder, in your editor, in grep. There's no gap between what the agent stored and what you can look at.
• Shared with you, and between agents. A memory service is generally scoped to an application and its users; the human's own writing is a different system. Here it's one corpus. I write into it by hand, and every connected client — Claude Desktop, Claude Code, Claude on the phone, an n8n workflow — reads and writes the same files on the same terms. A note I type on Sunday is context for an agent on Monday with no import step.
• Portability. The exit path from a folder of markdown is cp -r. No export format, no migration script, no question about what you'd be left holding if a project stopped being maintained. That's a property of files, not something this server does for you.
• Self-description. The rules live in the corpus rather than in client config. CLAUDE.md at the vault root tells every agent, on its first call, where things go and what frontmatter they carry, so conventions are versioned next to the notes they govern.

What the hosted shape buys you in exchange is real, and worth saying plainly. There's no Postgres to run, no pgvector version to keep current, no container to babysit — you get a memory layer by adding a dependency, which is a genuinely better trade for most people. And systems in that class typically do work this server deliberately doesn't attempt: pulling facts out of a conversation automatically, reconciling ones that contradict each other, and scoring relevance or decaying old memories so they stop crowding out new ones. Here an agent remembers something because it decided to write a note, and the judgment about what's worth keeping is the agent's, not the server's. The consolidation help planned here is a report that proposes and never writes (#359). If you want memory that curates itself, that's a fair reason to pick the other shape.

vs. an agent with raw file access

The other baseline isn't an MCP server at all: point Claude Code, a generic filesystem MCP, or any agent with file tools straight at the vault folder. That works — until a write goes wrong. An agent rewriting a whole file from its memory of an earlier read will eventually clobber a note, follow a symlink somewhere it shouldn't, or "tidy up" your .obsidian config. Nothing in a raw file API pushes back. This server's write path is shaped by exactly that kind of incident, and it assumes the caller will eventually do something wrong:

• Targeted edits instead of rewrites. edit_note can address a find-string or a single section rather than replacing the file, and dry_run=True returns the unified diff before anything lands. set_frontmatter mutates YAML structurally and leaves the body byte-identical.
• No-clobber defaults. create_note and write_file refuse to overwrite an existing file; replacing one is an explicit opt-in.
• Atomic writes. Content is staged and renamed into place against a descriptor opened at validation time — a note is never left half-written, and the file that gets replaced is the file that was checked.
• Reversible deletes. delete_note and delete_file soft-delete into .trash/ with a non-replacing rename; permanent=True is the explicit escape hatch, not the default.
• Kernel-proved containment. Paths resolve under the vault root via openat2(RESOLVE_BENEATH | RESOLVE_NO_SYMLINKS | RESOLVE_NO_MAGICLINKS), writes refuse a symlink as the final component, and dot-directories (.obsidian, .git, .trash) are out of reach of every tool.
• Bounded responses. Reads are capped and truncation is data (truncated, next_offset, an outline) rather than silent loss, so one huge note can't flood an agent's context into a bad edit.
• An audit trail. Every call is attributed to a key and logged; the control panel shows who touched what, and when.

When an agent misbehaves through this server you get a refused call, a diff, a trash entry, and a usage-log line. When it misbehaves with raw file access you get whatever git diff can recover — if the vault was in git at all.

Who this is for

• Homelab folks who already run Postgres and Docker, or are happy to spin them up. The setup tax is the price of admission for the semantic and graph layers.
• People who keep an opinionated vault — task placement logic, frontmatter schemas, tag taxonomy — and want agents to follow those conventions on the first call instead of being briefed every session.
• Anyone running more than one MCP client (Claude Desktop, Claude Code, Claude in a browser, n8n) against the same notes and tired of re-explaining the vault to each.
• Folks who want agent memory to live as plain markdown files they can read, edit, grep, and version-control, not in an opaque vector store or a managed memory service.

Who this isn't for

• "I just want Claude to read my notes" with the lightest possible setup. Use one of the desktop-local tools above; you don't need this.
• Obsidian is open on the machine the agent runs on and you need no remote access. The official CLI or the Local REST API plugin's MCP server is the lighter choice.
• Anyone unwilling to run a database. There is no SQLite fallback; pgvector is doing real work, and a managed Postgres with pgvector support is part of the stack.
• People who want a turnkey hosted product. This is a self-hosted server you run yourself.

Control panel

The server ships with a built-in admin UI for the parts of operations that are easier to look at than to query: minting keys, watching the indexer, eyeballing tool-call traffic, and resetting embeddings when you switch providers.

Usage

Per-tool-call audit log with a 14-day request histogram. Every MCP call is recorded with the calling key, tool name, duration, and response size — useful for noticing a misbehaving agent burning tokens on something it shouldn't.

[Usage]

API keys and OAuth clients

Bearer keys with read / readwrite scopes for API clients, and a separate OAuth 2.0 PKCE flow for clients like ChatGPT, Claude Desktop, and claude.ai that expect a proper authorization-code dance. The OAuth server supports public (none) and confidential (client_secret_post) token-endpoint authentication plus refresh tokens.

Clients identify themselves in one of two ways:

• A Client ID Metadata Document (CIMD, the MCP 2026-07-28 default). The client's client_id is an https URL to a JSON document it publishes, and the server fetches it through the same hardened, SSRF-guarded client as import_from_url. Claude (claude.ai, Desktop, mobile), Claude Code and ChatGPT use this whenever the server advertises it, which it does by default. Claude Code's loopback callback (http://localhost:<any port>/callback) works this way.
• Dynamic Client Registration (/register, RFC 7591, deprecated by MCP but kept) remains the fallback for any client that sends no URL.

The consent page names a CIMD client by the host its document was fetched from, never by the name the document gives itself. Every authorization response carries the RFC 9207 iss parameter.

Each client's page lists its grants — one row per /authorize approval, not per token — with a Revoke control and a permission select per grant, so revoking really ends the session instead of leaving a refresh token to mint a replacement. Revoked and expired rows stay listed, dimmed, for a week.

[API keys] [OAuth clients]

Vault browser

A read-only file tree of the mounted vault, mostly for sanity-checking that the container sees what you think it sees.

[Vault]

Settings

Indexer status, current embedding provider and model, vault path, and the danger zone: Reset embeddings (drops and recreates the embeddings column at the configured dimension — use it when switching providers) and Force re-embed (keeps the column, clears every note's embedded-content hash so the next pass re-embeds the vault). Both pause the indexer while they run.

The dashboard separates two things that used to be conflated: Last run is the indexer's own heartbeat — the last pass that completed, whether or not anything had changed — and Last change detected is the newest indexed_at on any note. A quiet vault makes the second one old while the indexer is perfectly healthy.

[Settings]

Quick start

Deploying on a VPS from scratch? See DEPLOYMENT.md for the full walkthrough: Postgres setup, Caddy and TLS, vault sync via Nextcloud, and the gotchas that bite first-time deploys. Running Kubernetes? See docs/deployment-kubernetes.md and the kustomize manifests in deploy/kubernetes/.

The bundled Caddy configuration fails closed on /admin, /api, and /authorize; replace its placeholder basic-auth hash before starting it.

Prerequisites

• Docker and Docker Compose
• A PostgreSQL 16 instance reachable from the container, with pgvector 0.8.0 or newer installed
• Either an Ollama instance running an embedding model (bge-m3, the default, or qwen3-embedding:0.6b, which the reference deployment runs; see Embedding model choice), or an OpenAI API key. Anything that speaks the OpenAI embeddings protocol works (Azure OpenAI, OpenRouter, Together, etc.).
• Linux, kernel 5.6 or newer (see below)

System requirements

The server checks these at startup and tells you which one failed rather than misbehaving later.

Linux kernel ≥ 5.6. Every directory below the vault root is opened with a single openat2(RESOLVE_BENEATH | RESOLVE_NO_SYMLINKS | RESOLVE_NO_MAGICLINKS), which is what makes the kernel — not the application — prove that a write stayed inside the vault. There is no fallback: on an older kernel, or under a container seccomp profile that blocks openat2, the server logs the reason and exits non-zero.

Kernel ≥ 5.8 for file transfer. statx()'s STATX_MNT_ID is how a publication refuses a destination that sits on a different mount than the staging directory (a nested bind mount under the vault root would otherwise fail only after a whole upload body had streamed). Below 5.8 the server logs one warning and starts: request_upload, import_from_url and PUT /transfer/upload refuse, and everything else — reads, note writes, search, downloads, the panel, OAuth — is unaffected. /health reports it as transfer_mount_check_available.

pgvector ≥ 0.8.0. Filtered semantic search needs hnsw.iterative_scan, which landed in 0.8.0. An older extension accepts the setting as an unknown placeholder and silently runs a plan that drops post-filter candidates — silently worse search results — so the server exits instead. Fix with ALTER EXTENSION vector UPDATE or a newer database image.

Filesystem. Case-sensitive and non-normalising (ext4, xfs, and the usual bind mounts). It must support hard links within the vault root and renameat2(RENAME_NOREPLACE); without those, note creation, move_note and the soft delete refuse with a named error rather than degrading to a publish that can clobber. O_TMPFILE is wanted but optional: where it is unavailable, set VAULT_ALLOW_NAMED_STAGING_FALLBACK=true to accept named staging instead (see Configuration). macOS and Windows hosts are out of scope; run the container on a Linux VM.

1. Clone, configure, point at your vault

git clone https://github.com/maxkuminov/obsidian-mcp.gitcd obsidian-mcpcp .env.example .env$EDITOR .env

Using a bundled stack (docker-compose.simple.yml or docker-compose.proxy.yml)? It runs PostgreSQL with two identities: set OBSIDIAN_DB_PASSWORD in .env (the app's non-superuser role, from which the compose file builds DATABASE_URL) and copy postgres.env.example to postgres.env with its own POSTGRES_PASSWORD (the database superuser, which only the postgres container sees). make init generates both. See DEPLOYMENT.md.

In docker-compose.yml, point the /obsidian volume at your vault:

volumes:  - /path/to/your/vault:/obsidian

2. Pick an embedding backend

Option A, OpenAI (zero local infra):

EMBEDDING_PROVIDER=openaiOPENAI_API_KEY=sk-...EMBEDDING_DIMENSIONS=1024OPENAI_EMBEDDING_MODEL=text-embedding-3-small

The server validates OPENAI_API_KEY at startup and refuses to boot if it's missing.

Option B, Ollama (self-hosted, GPU recommended):

EMBEDDING_PROVIDER=ollamaOLLAMA_URL=http://your-ollama-host:11434EMBEDDING_ALLOW_PLAINTEXT=trueEMBEDDING_MODEL=bge-m3EMBEDDING_DIMENSIONS=1024

This is the default. Omitting EMBEDDING_PROVIDER falls back to Ollama. On a new install, consider qwen3-embedding:0.6b (also 1024 dimensions) with a query instruction, which is what the reference deployment runs; read Embedding model choice before the first index pass, because changing the model later means a full re-embed.

The embedding URL must be https, or http to a loopback host (localhost, 127.x, ::1). Plaintext http to any other host — another container such as http://ollama:11434 included — refuses to start unless EMBEDDING_ALLOW_PLAINTEXT=true acknowledges that chunks and queries cross that hop unencrypted. .env.example ships with it set for that reason; drop it once the endpoint is https (use EMBEDDING_CA_FILE for an internal CA). Inside a container, localhost is the container itself, so an Ollama on the Docker host still needs the override.

3. Deploy

make init       # data dirs and .env from template (skip if you've already edited)make db-init    # create database, user, and pgvector extensionmake deploy     # build, push to local registry, run migrations, recreate container

The first deploy backfills the index, the wikilink graph, and the embeddings. For a 2 to 3k-note vault on Ollama with a GPU this takes a few minutes. On text-embedding-3-small it's seconds.

Prebuilt images are public at ghcr.io/maxkuminov/obsidian-mcp (:v<release> per release tag, :main for the tip), built and scanned by CI; a release tag is the image CI proved for that commit, retagged, never rebuilt. docker pull ghcr.io/maxkuminov/obsidian-mcp:v<release> needs no token. Or build locally as above.

4. Connect a client

Mint an API key in the control panel, then point your MCP client at:

URL:  https://obsidian-mcp.<your-domain>/mcpAuth: Bearer omcp_...

For Claude Desktop, add to claude_desktop_config.json:

{  "mcpServers": {    "obsidian": {      "url": "https://obsidian-mcp.<your-domain>/mcp",      "headers": { "Authorization": "Bearer omcp_..." }    }  }}

For Claude Code:

claude mcp add obsidian --transport http \  --url "https://obsidian-mcp.<your-domain>/mcp" \  --header "Authorization: Bearer omcp_..."

The first thing any agent should do in a new session is call get_vault_guide(). That's how it learns your folder structure, naming conventions, and YAML schema before it writes anything.

Versioning and compatibility

From 1.0 the version number is a promise about the MCP tool contract: the 28 tools' names, their argument schemas, the shape of their result lines, and the MCP-REFUSAL codes. An agent written against 1.0 keeps working against every 1.x.

• The wire form is pinned. tests/fixtures/tools_list_full_v1.json is the complete tools/list result an API key or a readwrite OAuth grant sees (every field of all 28 tools), and tests/fixtures/tools_list_readonly_oauth.json pins the 19 names a read-only OAuth grant lists. CI fails on any drift; a deliberate change regenerates the fixture and says so in its spec. Result line shapes are specified in openspec/specs/ and pinned by each tool's tests, not by one fixture.
• Refusal codes are a closed set. Every code an MCP-REFUSAL line can carry is declared in src/services/refusals.py (CODES), and a refusal with any other code cannot be constructed. That module is the catalogue; there is no separate list. The transport 429s (failed authentication, transfer redemption, body_memory) are outside that line by design.
• Within 1.x, additions only: new tools, new optional arguments whose default leaves the old result byte-identical, new header tokens or trailing fields, new refusal codes. Never a rename, a removal, a changed default that alters an existing result, or a changed meaning of an existing field or code. A change of that kind is 2.0, announced one minor release earlier in the release notes. Unknown arguments are refused, so a client that sends a newer argument to an older server gets a refusal, not a silently ignored filter.
• Settings. A new setting ships with a default that keeps the previous behaviour. Every setting change is in the release's upgrade notes, in the Configuration table and in .env.example.
• Schema. Every release names its migrations and says whether they refuse startup or need a rebuild or re-embed, and DEPLOYMENT.md documents each one's downgrade (it does so for every migration from 033 on).

Not part of the contract: the control panel and its /api/* endpoints, log fields, the search-evaluation harness (scripts/search_eval/), the internal MCP SDK version, and ranking quality. Which notes come back, and in what order, moves with the index, the embedding model and the reranker; the harness measures it (docs/architecture/search.md). MCP_SANDBOX_MODE, the registry-evaluation switch, is outside the promise too.

Upgrading

Pull, then make deploy (or rebuild your compose stack); migrations run on start.

Across the OIDC account-linking release (migration 037, #416):

• Migration 037 adds six nullable link columns and five constraints to users, with no backfill, and takes seconds. With SSO off nothing else changes. Its downgrade drops every link.
• OIDC_SUBJECT_USER_IDS is retired. A deployment that still sets it refuses to start. If yours did, do this in order (it is also in the startup message): (1) on the old release, confirm a local admin password works; (2) set LOCAL_LOGIN_ENABLED=true; (3) copy the map's subject-to-user pairs, then remove the OIDC_SUBJECT_USER_IDS line (delete it; an empty OIDC_SUBJECT_USER_IDS= fails to parse at startup); (4) deploy and start; (5) sign in with the password and, on each user's edit page, Single sign-on → Bind subject with their former subject; (6) verify an SSO sign-in per user and one MCP client's /authorize consent; (7) optionally set LOCAL_LOGIN_ENABLED=false again, once the startup log shows no CRITICAL lockout line. Details: docs/native-oidc.md, "Upgrading from OIDC_SUBJECT_USER_IDS".
• Wider provider profile. RS256, PS256 and ES256 ID tokens (as advertised by discovery), client_secret_post via OIDC_TOKEN_AUTH_METHOD, OIDC_EXTRA_SCOPES, OIDC_DISPLAY_NAME and OIDC_CA_FILE are new. OIDC_REQUIRED_GROUP is now optional; set, it works as before. The requested scopes become openid email profile. A provider that advertises RFC 9207 must now return iss on the callback.
• SSO logins now stamp the user's last-login time.

Across the reranker-documents release (#402):

• No migration, nothing re-derived or re-embedded. rerank=true now sends each candidate note's embedding input (title, heading path, tags, then the chunk) for its cosine-best chunk and up to three more of its already-fetched chunks, at most RERANK_MAX_DOCUMENTS per call (new, default 100, 50–200; 50 restores the previous count); a note scores the best of its documents. rerank=false is byte-identical. The usage row gains rerank_documents.
• RERANK_TIMEOUT_MS accepts 100–30000; the default stays 1,500. Size it with the document budget: RERANK_TIMEOUT_MS ≥ RERANK_MAX_DOCUMENTS × average characters × per-character cost, doubled when the GPU is shared with the embedding model. The reference deployment runs 20000 (DEPLOYMENT.md, "Reranker (optional)"). Under MCP_CONCURRENCY_MODE queue or enforce, keep it at or below MCP_CONCURRENCY_WAIT_SECONDS × 1000 unless make concurrency-report passes with rerank=true traffic.
• Rollback is a redeploy of the previous image. Set RERANK_TIMEOUT_MS to 10000 or less first: the previous image's range ends there and refuses to start above it.

Across the Client ID Metadata Documents release (migration 036, #357):

• Migration 036 adds one nullable column and one CHECK to oauth_clients, with no backfill, and takes seconds. Its downgrade refuses while CIMD clients exist (delete them in the panel first).
• Connectors switch to CIMD on their next connection. Existing DCR grants keep working. claude.ai, Claude Code and ChatGPT stop calling /register. OAUTH_CIMD_ENABLED=false is the rollback lever.
• Every authorization redirect now carries iss (RFC 9207). A client that compares it does so against the metadata issuer, which is unchanged.
• resource (RFC 8707) is logged, not enforced. Each /authorize and /token request is classified (resource_class) in the security log. Refusing a foreign resource is a later change, gated on 14 days of these logs (DEPLOYMENT.md).
• The consent page's CSP gains form-action http://localhost:* http://127.0.0.1:* for native loopback callbacks.

Across the scoped tools/list release (#358):

• No migration and no setting. An OAuth grant without readwrite now lists 19 tools: the nine write-required tools (create_note, edit_note, move_note, delete_note, set_frontmatter, write_file, delete_file, import_from_url, request_upload) are left out. API keys of either permission and readwrite grants still list all 28, byte-identical to before. A read-only client that cached the 28 and calls a hidden tool gets permission_denied, as before.
• No list_changed is sent. After a grant's scope is changed in the panel, the connector sees the new list when it reconnects.
• One refusal changes for read-only callers. The refusals write_file / delete_file made before their permission check (a malformed expected_hash, no_incumbent for write_file(overwrite=false)) are now permission_denied; a readwrite caller's refusals are unchanged.

Across the keyword-tags release (migration 035, #391):

• Tags become searchable at once. Migration 035 adds each note's tags to its keyword vector in place, in SQL, under the index generation lock (tens of seconds on a few thousand notes); nothing is re-embedded or re-derived and no setting is new. make deploy stops the app before the migration; on a hand-driven compose stack: stop → alembic upgrade head → start.
• Run the read-only check first. 035 aborts with nothing written if a tagged vector would cross PostgreSQL's 1 MiB tsvector limit; under Kubernetes Recreate that is downtime until the digest is reverted. DEPLOYMENT.md, "Upgrading across migration 035" has the statement to run against the live database before merging and the remedy.
• Rollback: alembic downgrade 034 with the new image strips the tags part, then deploy the previous image.

Across the chunk-minimum release (CHUNKING_VERSION 2, #390):

• No migration and no new setting, but a second gradual re-embed. Chunks shorter than 200 characters are no longer stored (a short section is grouped with its neighbours), windows are cut at a line or word start, and long whitespace runs are sent collapsed. Startup does not refuse; every note is selected again in the background under the per-tenant embed budgets.
• Until your notes drain, the index_transition: N notes not yet re-chunked; … line returns on semantic_search, find_related and hybrid_search, and /health reads indexer.rechunk: "in_progress", then "complete". A section that did not change keeps its vector; each pass logs Embedding complete: N notes, X chunks (S submitted, R reused).
• Rollback is a redeploy of the previous image, which re-embeds every note stamped 2 in full (DEPLOYMENT.md).

Across the search-confidence release (#389):

• No migration. Every semantic_search and hybrid_search page gains one header token: — confidence: ok (at the default threshold 0 that is the only byte that changes). A parser that matched the old header exactly needs to allow it. Usage rows gain confidence, top_similarity and (hybrid) keyword_all. One optional setting, SEARCH_LOW_CONFIDENCE_SIMILARITY (default 0 = off); leave it at 0 until the calibration recipe in DEPLOYMENT.md passes for both tools.
• The search-eval harness reports recall@20 and recall@50, runs at --depth 50 by default and writes report.json schema_version 2; a runfile: baseline staged at depth 20 scores a null recall@50, so re-run it for a delta. It gains a confidence subcommand.
• Rollback is a redeploy of the previous image; the new usage keys stay and are ignored.

Across the reranker release (#367):

• No migration, nothing re-derived or re-embedded. semantic_search and hybrid_search gain an optional rerank argument (default false, output unchanged), and four optional settings (RERANK_URL, RERANK_MODEL, RERANK_TIMEOUT_MS, RERANK_ALLOW_PLAINTEXT). Nothing changes until you set RERANK_URL. The concurrency class mapping is unchanged, so readiness evidence carries over. A reference setup is in Reranker reference setup.
• Rollback is a redeploy of the previous image. A client that keeps sending rerank is then refused by the unknown-argument rule, which is the normal cost of rolling back a new argument.

Across the hybrid-search release (#354):

• No migration, nothing re-derived or re-embedded. A new tool, hybrid_search, and one optional setting, HYBRID_RRF_K (default 60).
• A new concurrency-readiness epoch. hybrid_search shares the embedding class with semantic_search, so CLASS_MAPPING_VERSION moves to 5 and make concurrency-report evidence gathered before the upgrade does not carry over.
• Rollback is a redeploy of the previous image; the tool's usage rows stay and are ignored.

Across the change-feed release (migration 033, #351):

• One migration, no re-derive. 033 adds notes_metadata.content_changed_at (backfilled from each note's modification time) and three change-feed tables. Nothing is re-parsed or re-embedded. The feed starts with the first pass after the upgrade; nothing before 033 has events.
• The feed tables are durable. make db-backup and the Kubernetes backup Job dump their rows (history, cursors and retention watermarks cannot be rebuilt from the vault). A restore rolls the feed back to the dump, so a cursor taken after it is refused and the agent restarts from a date.
• A new concurrency-readiness epoch. get_changes joins the scan class, so CLASS_MAPPING_VERSION moves to 4 and make concurrency-report evidence gathered before the upgrade does not carry over.
• Four new settings, all with working defaults: CHANGE_FEED_RETENTION_DAYS, CHANGE_FEED_MAX_EVENTS_PER_SCOPE, CHANGE_FEED_MAX_EVENTS_PER_PASS, CHANGE_FEED_LOCK_TIMEOUT_MS (see Configuration).
• Rollback: the previous image neither reads nor writes any of these objects; alembic downgrade 032 drops them and loses only the feed's history.

Across the task-listing release (migration 032, #350):

• One full re-derive pass after the upgrade, and no re-embedding. The extraction version moves to 4, so the first pass re-parses every note once and stores its checkbox tasks for the new list_tasks tool — the same cost as the earlier extraction-version bumps. Version 4 shares version 3's embedding cleaner, so nothing is re-embedded and no provider call is made. Until that pass finishes, list_tasks reports the notes it has not scanned yet in its header, and move_note(rewrite_links=True) refuses with index_not_ready as it does after every bump.
• A new concurrency-readiness epoch. list_tasks joins the scan class, so CLASS_MAPPING_VERSION moves to 3 and make concurrency-report evidence gathered before the upgrade does not carry over.
• Rollback is roll-forward. A bare redeploy of the previous image re-extracts every note once, and because it has no version-4 cleaner it also clears every note's embedding certification; the embed pass then re-chunks every note and reuses every stored vector (no provider calls), with semantic_search marking those notes stale until it catches up. Rolling forward again is one more re-derive pass.
• No new setting.

Read this first when upgrading across the keyword-ranking release (migration 031, #344):

• Keyword ranking changes at once; snippets need one rebuild. Migration 031 rewrites every keyword vector in place so titles and aliases rank above bodies — no operator step, and make deploy now stops the app before a pending migration and recreates it after (on a hand-driven compose stack: stop → alembic upgrade head → start; Kubernetes already Recreates). Snippets come from an excerpt that only a write stores, so notes indexed before 031 show no snippet (the header counts them and startup logs the count) until they are edited or you run make rebuild-tsvectors once (on Kubernetes, deploy/kubernetes/base/rebuild-tsvectors-job.example.yaml with the app scaled to zero). No embeddings are touched, no API calls are made.
• New optional settings: RANK_WEIGHTS, KEYWORD_EXCERPT_CHARS, KEYWORD_SNIPPET_ROWS. See Configuration.

And across the internal-transport and panel-CSP release:

• Breaking: the server refuses to run as a PostgreSQL superuser (#324). A Compose install created from docker-compose.simple.yml or docker-compose.proxy.yml before this release connects as the cluster superuser and must run the one-time conversion in DEPLOYMENT.md, "Upgrading: split database roles" (a backup, a new postgres.env, OBSIDIAN_DB_PASSWORD in .env, one script); the new compose files will not serve on an unconverted volume. Any other deployment whose DATABASE_URL names a superuser is refused at startup with a message; switch it to a non-superuser owner of the database, or set DATABASE_ALLOW_SUPERUSER=true as a stop-gap. The Kubernetes bundle is unaffected. A DATABASE_URL password of changeme / CHANGE_ME is now refused, and the built-in default URL carries no password.
• Breaking: plaintext embedding endpoints must be acknowledged. If the active embedding URL (OLLAMA_URL, or OPENAI_BASE_URL with the OpenAI provider) is http:// to a non-loopback host — the http://ollama:11434 default included — add EMBEDDING_ALLOW_PLAINTEXT=true to .env before deploying, or the server refuses to start with a message naming the setting.
• Database TLS has one source. A TLS parameter in DATABASE_URL (?ssl=…, ?sslmode=…) or any PGSSL* environment variable is refused at startup; move it to DATABASE_SSL_MODE. The default, prefer, is the behaviour you had before.
• Embedding clients ignore the environment's network settings. HTTP(S)_PROXY, SSL_CERT_FILE / SSL_CERT_DIR and .netrc no longer apply to the embedding hop. Use EMBEDDING_CA_FILE for an internal CA.
• The panel now sends a nonce-based Content-Security-Policy (PANEL_CSP=enforce), and htmx is gone from it. If a panel control misbehaves, set PANEL_CSP=report-only (or off) and recreate the container; no rebuild.
• New optional settings: DATABASE_SSL_MODE, DATABASE_SSL_CA_FILE, DATABASE_SSL_CERT_FILE, DATABASE_SSL_KEY_FILE, EMBEDDING_ALLOW_PLAINTEXT, EMBEDDING_CA_FILE, PANEL_CSP. See Configuration, and DEPLOYMENT.md for moving both hops to verified TLS.

Cost expectations

If you go the OpenAI route (the realistic path on a CPU-only VPS), the first-index spend is small and the steady state is nearly free. Rough numbers assuming an average note around 1,500 tokens (three 512-token chunks), at OpenAI's published rate at time of writing:

After the first index, only changed notes are re-embedded. Ongoing cost is proportional to edits — pennies a month for a typical vault.

If you self-host Ollama with a GPU, embedding cost is whatever your power bill is. Ollama on CPU works but is too slow to be usable on a vault of more than a few hundred notes.

The self-describing vault

This is the part most "MCP for Obsidian" projects miss. They stop at read, write, and list. The interesting question isn't "can the agent reach the files," it's "does the agent know the rules?"

If you have an opinionated vault — task placement logic, folder conventions, required frontmatter, tag taxonomy — an agent with write access can do real damage without that context. Tasks land in the wrong folder. Bare-date filenames collide with templates. Wrong tags break Dataview queries. The data layer works fine; the context layer is where the failures show up.

The fix is small. Keep a machine-readable instruction file (CLAUDE.md at the vault root) that describes the system's own rules. Expose it as a dedicated tool. Every connecting agent calls it once at the start of a session and immediately knows how the vault works. Update the file, every agent sees the change on the next call. No client-side config. No system-prompt injection. The vault is authoritative about its own rules.

get_vault_guide() does exactly this. It returns a generic Obsidian primer (wikilink syntax, embed syntax, tag conventions, common plugin literals) plus the vault's CLAUDE.md live. The write-tool descriptions and the server's instructions both point at it — as a description of where the vault's conventions live, not as an order — so the agent gets pulled into the right behavior even without prompting.

Multi-user mode

Single-user mode is the default and works exactly as described above — one vault, one set of API keys, no in-app user concept. Multi-user mode is an opt-in flag that turns the same container into a small multi-tenant deployment: in-app username/password login, per-user vault scoping, an admin role for troubleshooting, and a regular-user role that sees only its own keys/OAuth clients/usage. One container, one Postgres, strict isolation between users.

Enable it on an existing deployment with no data loss — your current vault and keys carry over to the bootstrap admin.

Enabling

1. Set MULTI_USER_MODE=true and a strong SECRET_KEY in .env (openssl rand -hex 32 is fine). The app refuses to start with a placeholder SECRET_KEY unconditionally — single-user mode included — so this is not something the flag turns on.
2. make deploy (or docker compose up -d --force-recreate).
3. Visit the panel. Because the users table is empty, you're routed to /admin/register — the one-time bootstrap form. It's still behind Traefik's chain-oauth@file middleware, so only people Traefik already trusts can claim admin.
4. Register with a chosen username and password. The bootstrap form pre-fills vault_path with whatever VAULT_PATH was set to, so your existing notes immediately belong to this new admin. No re-index, no re-embed, no data loss — every previously indexed note, API key, OAuth client, and usage log row gets backfilled to the bootstrap user in a single transaction.

Inviting users

1. Edit docker-compose.yml to add a volume mount for the new user's vault under /vaults/<username>. Host paths with spaces must be quoted as a single YAML string:

   yaml

   volumes:  - "/storage/vaults/alice:/vaults/alice"  - "/storage/shared/bob/Obsidian:/vaults/bob"

   make deploy to apply.

2. In the panel, /admin/users/create — pick a username and set an initial password.

3. /admin/users/{id}/edit — set the user's vault_path to the container path you just mounted (e.g. /vaults/bob). The form shows a dropdown of unassigned /vaults/* directories that exist on disk.

4. Share the credentials out-of-band. The user logs in at /admin/auth/login, gets their own keys/OAuth/usage views, and cannot see other users' notes.

What admins see

Admins see API keys, OAuth clients, and usage logs for all users; they own the Settings page (embedding provider, indexer trigger, danger zone) and the Users page. Admins do not browse other users' vault contents through the panel — that's intentional. Troubleshooting another user's vault means either inspecting it via docker exec or temporarily reassigning their vault_path, not UI snooping.

Rolling back

Set MULTI_USER_MODE=false, restart. Existing API keys keep working (per-user filters skip when no user context is set), the login UI and session cookies disappear, and the panel falls back to its Traefik-OAuth-only mode. The schema stays in place, so flipping back to multi-user later resumes where you left off without re-bootstrapping (the users table is non-empty, so /admin/register is closed).

Single sign-on (optional)

Multi-user deployments can add "Sign in with SSO" through any standards-compliant OpenID Connect provider (Keycloak, Authentik, Authelia, Zitadel, Kanidm, Okta, Entra ID, …). It is off by default (OIDC_ENABLED=false), and with it off nothing changes.

• It signs in existing accounts only. An admin links each account on its edit page, Single sign-on card: Expect email (the first sign-in within 7 days whose provider-verified email matches links the account, once) or Bind subject (paste the provider's sub, linked immediately). After that only the provider's issuer and subject identify the user; a renamed or re-addressed provider account changes nothing. An identity that is not linked is refused with a page that does not say why.
• No provisioning, no role mapping. Signing in never creates a user, changes the admin flag or assigns a vault, and /admin/register is closed while SSO is on. Create users first.
• The server stays the OAuth server for MCP clients. SSO only changes how the panel session is made; that session also covers the consent page, so connecting a client can go through SSO. API keys, grants and tokens are unchanged.
• Password login stays unless you set LOCAL_LOGIN_ENABLED=false. While it is off, the panel refuses any change that would leave no active admin linked to SSO, and startup logs a CRITICAL line if there is none. Re-enabling it and restarting is the recovery path.
• Unlinking clears the link and signs that user out everywhere. Removing someone at the provider stops new logins but does not end their existing sessions, keys or grants: deactivate the local account for that.

docs/native-oidc.md has the provider requirements, every setting, a click-by-click Keycloak walkthrough, the setup order, break-glass recovery and troubleshooting by refusal reason.

Constraints and known limits

• The indexer iterates active users sequentially each cycle. Fine for tens of users; hundreds would need parallelization.
• Password recovery is admin-driven — there's no email-based reset. A signed-in user can rotate their own password at /admin/account (current password, new password, confirmation; minimum 12 characters), which signs their other browsers out and keeps the one they changed it from signed in. The admin reset stays the recovery path for somebody who cannot sign in at all, and it also ends every live session of the account it resets.
• /admin/auth/login and /admin/account/password are rate-limited at 5 requests per minute; the login limit is keyed on the client address, and the password change carries two independent limits — one per account, one per address. The limiter's storage is in-memory and per-process, so counters reset on restart. The Traefik OAuth gate in front of the panel is still the main brute-force defense; if you expose /admin/auth/login to the open internet, put a rate-limit middleware in front of it as well.
• Panel sessions are server-side rows (user_sessions), so logging out, changing a password, a deactivation or a delete really ends them. The trade-off: the first deploy of the build that introduced the registry signs every live panel session out once, because a cookie issued before it carries no session id and is refused rather than grandfathered. Everyone signs in again; nothing else changes.
• The vault_path validator does not resolve symlinks, so an admin can technically point a user at host files via a symlinked /vaults/<name>. Treat /vaults/ as an admin-trust boundary. What is checked, since the vault-root overlap guard: two active users' roots may not name overlapping directories. Each root is opened once and compared by inode identity — (st_dev, st_ino), which catches a symlink alias or a bind mount naming one directory twice — and by a component-wise containment test over the two canonical real paths in both directions, which catches an ancestor/descendant pair like /vaults/team and /vaults/team/private. A conflicting assignment is refused in the panel naming the other user, and the same checks re-run before every index pass, so an alias created after the assignment quarantines both accounts: their MCP tools, index passes and transfer redemptions are refused until an administrator corrects it, and no index rows are deleted. A root that cannot be opened at all quarantines only its own account. What is still not detected, and the consequence: a bind mount that grafts one user's vault — or any mount nested inside it — to a path inside another user's root. mount --bind /vaults/b /vaults/a/inner leaves both root inodes distinct and both canonical paths outside each other, so neither check sees it, and user A can then read, overwrite and delete every note in user B's vault through the ordinary write tools, while A's index pass files B's notes under A's account so A's searches return B's content. The same gap covers an accessible alias of a root that could not be examined: that peer keeps serving. Neither condition is reported anywhere. Both require an administrator to write a bind mount into the deploy configuration — which is why /vaults/ and the compose file's mounts are the admin-trust boundary, not just the path strings. This is a permanent, stated limit rather than a pending fix: mount detection was specified, failed on a new topology in each of three review rounds, and was dropped. The operator rule: never mount one user's directory, or anything nested in it, inside another user's root.

Configuration

See .env.example for the full set with comments. For first-index spend on OpenAI, see Cost expectations above.

The MCP transport's request-body limit is derived, not configured: max(2 × MAX_FILE_WRITE_BYTES, 6 × 10 MB) + 1 MiB, which is 61 MiB with the defaults. It has to track the write caps so that every supported write is refused by the tool — with an actionable message — rather than by the transport with a bare HTTP 413. Raise MAX_FILE_WRITE_BYTES and the transport limit follows.

That limit is per request. The sum of request bodies in flight is bounded too, always (#322): each /mcp POST reserves its declared size against a budget derived from the container's memory limit (half of it, ÷ 8 for the parsing overhead — 128 MiB of raw body on a 2 GiB container), waits up to 15 s if it does not fit, and otherwise gets HTTP 429 with code: body_memory and Retry-After: 2. The server refuses to start in a container too small to admit one maximum-size body (about 1.1 GiB with the defaults). The MCP_BODY_* settings in .env.example only allow the safe direction.

Embedding model choice

Two local models are documented here. Both are multilingual and produce 1024-dimensional vectors, so EMBEDDING_DIMENSIONS=1024 fits either.

• bge-m3 is the code default (EMBEDDING_MODEL=bge-m3) and stays a fully supported choice. It needs no query instruction (EMBEDDING_QUERY_INSTRUCTION empty). The default is not changing: the stored-fingerprint guard refuses to start a deployment whose model changed, so a new default would stop every existing install that relies on it.

• qwen3-embedding:0.6b is what the reference deployment has run since 2026-10-31, with a query instruction. The instruction is applied to queries only (notes are embedded without one) and is outside the stored fingerprint:

  env

  EMBEDDING_PROVIDER=ollamaEMBEDDING_MODEL=qwen3-embedding:0.6bEMBEDDING_DIMENSIONS=1024EMBEDDING_QUERY_INSTRUCTION="Instruct: Given a question, retrieve the notes that answer it\nQuery: "

What was observed after the switch, on one user's vault (not a benchmark): cross-language recall between English and Russian queries and notes was strong, and hybrid_search became the day-to-day default. No harness A/B was run for the switch, so no quality gain is claimed. To measure it on your own vault, use the search-evaluation recipe in DEPLOYMENT.md.

Switching an existing install is a full re-embed (the procedure is below). For scale: on an 8 GiB Quadro P4000 shared by Ollama and the reranker, ~4,200 notes re-embedded in about 84 minutes. SEARCH_LOW_CONFIDENCE_SIMILARITY is model-dependent: its 0.50 candidate was measured on bge-m3, so re-calibrate after a switch.

Reranker reference setup

rerank=true is opt-in. Before #402 the reranker scored one heading-less chunk per note, which demoted some notes both search arms ranked first; it now scores each note on the embedded text (title, heading path, tags, chunk) of up to four of its fetched chunks (RERANK_MAX_DOCUMENTS). The #402 real-data gate ran on 2026-10-31 against the previous reranker documents and passed 4 of 6 reports (the owner waived one near-miss cell). Measured against rerank=false on the reference vault (2026-11-01, 196 queries), it improves semantic_search (overall recall@10 0.505 → 0.582, nDCG@10 0.389 → 0.442) and makes hybrid_search worse (MRR 0.533 → 0.465, nDCG@10 0.519 → 0.474), because plain hybrid is already the strongest ranking there. Keep rerank=false as the default, use rerank=true with semantic_search when the top few rows matter, and call hybrid_search without it. The intervals and the method are in DEPLOYMENT.md; measure on your own judgments after a model or reranker change.

Ollama has no rerank endpoint (ollama/ollama#3368), so the reference deployment runs llama.cpp's llama-server (image ghcr.io/ggml-org/llama.cpp:server-cuda12-*, pinned to a build) serving bge-reranker-v2-m3 as a Q8_0 GGUF with --reranking, behind TLS:

RERANK_URL=https://<host>:<port>/v1/rerankRERANK_TIMEOUT_MS=20000

Measured on the same P4000 with nothing else running: 50 documents of ~410 characters in 1.6 s, 50 documents of 2,000 characters in 7.3 s. Candidates are cut to 2,048 characters, so the 1,500 ms default deadline is too short for this card. A hybrid rerank=true call sends 67–68 documents on the reference vault and takes about 9–10 s uncontended (68 documents of 2,000 characters measured 9.1 s), so the reference runs 20,000 ms since 2026-11-01; the setting accepts 100–30000. Details, the sizing table and the startup checks are in DEPLOYMENT.md.

Switching providers or models

Different models produce vectors in different spaces, and cosine distance between two spaces is meaningless. So any change to what produced the stored vectors requires a full re-embed — not only a provider switch. That is every one of:

• EMBEDDING_PROVIDER
• EMBEDDING_MODEL (Ollama) or OPENAI_EMBEDDING_MODEL (OpenAI) — including a swap between two models of the same dimension, which the dimension guard cannot see
• EMBEDDING_DIMENSIONS
• CHUNK_SIZE and CHUNK_OVERLAP

Two things are deliberately not on that list. EMBEDDING_QUERY_INSTRUCTION touches only queries, so no stored vector depends on it. And the chunking format itself (heading-aware sections, the context prefix, and the 200-character minimum chunk length, which is a constant tied to the chunking version rather than a setting) is versioned per note: when it changes, the indexer re-embeds notes in the background over several passes, without a refused startup or a reset.

The server stores a fingerprint of that configuration and compares it at startup. On a mismatch it logs both fingerprints and the fields that differ, names the repair, and exits non-zero — so a model swap that used to mix two vector spaces in one column silently, for ever, now stops the process instead.

The steps, in this order:

1. Update .env.
2. make deploy (or docker compose up -d --force-recreate). The new container will refuse to start — at the fingerprint guard, or at the dimension guard if the width changed — and that refusal is the point: a container that will not start embeds nothing while the reset runs.
3. make reset-embeddings while it is down. The target is docker compose run --rm, so it starts a one-off container that reads your edited .env: it recreates the column at the new dimension, clears every embedded_content_hash, and records the new fingerprint in the same transaction.
4. Restart the service. It starts silently, because the stored rows really were produced under the configuration it is now running, and the next indexer pass re-embeds the vault.

This inverts the older reset-before-recreate advice. That ordering was safe only while nothing depended on a stored claim about the configuration; now the reset is what writes that claim, so it has to run with the new .env in place and with no old-configuration container able to embed against it. Skipping a step costs time rather than correctness — a database-level generation lock makes an old-configuration container's certifications refuse rather than land — but the ordering above is the one that never has to rely on it.

Maintenance waits for an in-flight index pass. That same generation lock is taken at the head of the index pass's transaction and held until it commits, so make reset-embeddings and make rebuild-tsvectors block until the pass finishes — up to a few minutes on a large vault — rather than interleaving with it. That wait is the required behaviour, not a stall to work around: a reset that landed mid-pass is precisely the interleaving that stores vectors from one configuration under a fingerprint naming another. Neither command sets a short lock timeout, and neither should be given one — and because the server sets a 60-second statement_timeout on every connection, both commands (and the panel's Danger-zone resets) lift that timeout for the acquisition itself and restore it once the lock is theirs. Without that, a command started against a live service was cancelled after a minute rather than waiting, which reads as a broken command instead of a busy index.

You can also use Settings → Danger zone → Reset embeddings in the control panel, which performs the same SQL — including the fingerprint record — while the server is running (pauses the indexer, runs the SQL, resumes).

The fingerprint records the configuration, not the model artifact. bge-m3 is a mutable Ollama tag (so is qwen3-embedding:0.6b), so ollama pull can replace the weights behind it, and OLLAMA_URL / OPENAI_BASE_URL are deliberately excluded from the fingerprint — repointing at another host or proxy is usually an infrastructure move that serves the identical artifact, and including it would demand a full re-embed for one. The consequence is an accepted limitation: replacing the artifact behind an unchanged model name — re-pulling a tag, or pointing at a host serving different weights under the same name — mixes vector spaces undetected. It requires make reset-embeddings, and no startup check will catch it if you skip that. No value available to the server distinguishes the two cases, and a probe would have to trust the endpoint it is checking.

Full-text search language(s)

keyword_search runs over a PostgreSQL tsvector. The text-search configuration it uses — the stemmer and stop-word dictionary — is controlled by FTS_CONFIGS. It defaults to english, which reproduces the historical behavior exactly, so existing deployments need no action.

The vector holds the note's title and frontmatter aliases at weight A, its tags at weight B (frontmatter tags and inline #tags; a nested tag also as each /-segment, so #ctx/phone is found by phone) and its body at weight D, per configured config, and ranking uses RANK_WEIGHTS — so a query that is a note's title ranks that note first, where it used to lose to a long note repeating the words, and a note tagged with a word ranks above one that only mentions it. Bare terms are ANDed; when the full matches do not fill the page, the same terms OR-ed fill the rest, tagged match: any. Each row carries a snippet: built from a stored excerpt of the indexed text (KEYWORD_EXCERPT_CHARS), computed for the first KEYWORD_SNIPPET_ROWS rows. The excerpt is written with the vector; make rebuild-tsvectors also stores it for notes indexed before it existed.

FTS_CONFIGS is a list, settable as JSON (FTS_CONFIGS=["simple","norwegian"]) or comma-separated (FTS_CONFIGS=simple,norwegian). Each note is indexed under every listed config, and a query matches if any listed config's parse hits. This is what makes a mixed-language vault work:

The setting is global — applied to every vault (consistent with EMBEDDING_MODEL, CHUNK_SIZE, etc., which are global too). For a mixed-language multi-user instance, set a superset (e.g. ["english","norwegian"], or ["simple"]). Per-user FTS config is a clean future extension but is not implemented.

A typo'd or uninstalled config name fails fast at startup with a message listing the configs available in your Postgres instance, rather than producing silent zero-result searches.

Changing FTS_CONFIGS requires a rebuild, and the server refuses to start until it has run. Stored tsvectors are computed at index time, so they go stale when the config list changes — and a stale stemmer is not merely incomplete. Under english the token running is stored as the lexeme run, so a query under simple for run matches a note that does not contain the word — a false positive, indistinguishable from a real hit. Keyword vectors therefore fail closed exactly as embeddings do: the server stores a fingerprint of FTS_CONFIGS, compares it at startup, and on a membership change logs both lists and the differing entries, names the rebuild, and exits non-zero. (Reordering the same names is not a change: a note is indexed under every config and a query matches if any hits, so order changes nothing and is not compared.)

The runbook:

1. Edit FTS_CONFIGS in .env.
2. make deploy. The new container refuses at the keyword fingerprint guard and stays down.
3. make rebuild-tsvectors. It rebuilds every scope that holds rows — every owner, including rows with no owner in single-user mode — in one transaction, and records the new fingerprint only if every one of them reported a completed rebuild. It is all-or-nothing: one scope it cannot rebuild rolls the whole thing back, names the scope and the reason, and writes no fingerprint, because the fingerprint is a single claim about every retained row.
4. Restart. It starts silently.

If step 3 names a scope it could not rebuild — a user whose vault is not assigned, a tenant still re-deriving its provenance, or ownerless rows under multi-user mode — there are three recourses, in order of preference:

• Settle the scope: assign or delete the user, or let the re-derive finish, then re-run the rebuild.
• Delete or reassign the ownerless rows, then re-run the rebuild.
• Put FTS_CONFIGS back to its previous value. That clears the refusal immediately, with no rebuild at all — a configuration edit is always reversible, which is what keeps this refusal from being an outage.

The rebuild re-reads each note and recomputes its content_tsvector under the new config(s). It rebuilds the keyword index only — it does not touch embeddings/vectors and makes no API calls, so it finishes in seconds for a few thousand notes. (Do not confuse it with the expensive make reset-embeddings flow.)

Tokenization caveat: the tsvector parser still splits on punctuation and hyphens regardless of config, so bge-m3 tokenizes to bge + m3. simple preserves word forms, not punctuation-bearing strings; exact-string-with-punctuation matching would need a trigram index and is out of scope.

Response size limits

A tool result is model input. Whatever read_note returns is fed straight back into the caller's next request, so an unbounded read is an unbounded prompt — and the caller usually finds out only when its inference provider rejects the request.

MAX_READ_RESPONSE_CHARS (default 40,000, roughly 10K tokens) bounds what read_note and the text results of read_file return. It is a different limit from MAX_FILE_READ_BYTES, which bounds what the server reads off disk. A 3 MB note is comfortably within the 10 MB read cap and will still destroy a context window; both caps are needed and they have different correct values.

It applies per component, not once to the whole response: the content window gets the cap, the heading outline gets it independently, and the metadata fields (title, tags, frontmatter_yaml and its JSON view, heading) share a third. A truncated read can carry all three, so budget for a worst case of roughly 3 × MAX_READ_RESPONSE_CHARS plus fixed prose — doubled again because the MCP result carries both structured content and a JSON text block, and multiplied by JSON escaping for content that is mostly control characters.

When a note exceeds the cap you get the first window plus truncation as data — truncated, the next_offset to continue from, total_chars — and, for a whole-note read, an outline of the note's sections:

{"entries": [  {"ordinal": 1, "depth": 1, "text": "Client Records",   "size": 2855343, "exceeds_cap": true,  "duplicate": false},  {"ordinal": 2, "depth": 2, "text": "Balance Sheet.xlsx",   "size": 391199,  "exceeds_cap": true,  "duplicate": false},  {"ordinal": 3, "depth": 2, "text": "Lease Agreement.pdf",   "size": 464,     "exceeds_cap": false, "duplicate": false},  {"ordinal": 4, "depth": 2, "text": "Invoice 2025-044.pdf",   "size": 1075,    "exceeds_cap": false, "duplicate": true} ], "truncated": false}

Paging a multi-megabyte note 40K at a time is technically possible and practically useless, so prefer the outline: read the one section you want with read_note(path, section="Lease Agreement.pdf"). Sections are addressable three ways — the #N ordinal shown in the outline, the Parent/Child path-style form, and exact heading text. The ordinal is the only form that separates duplicate sibling headings, which share every ancestor and so cannot be disambiguated by path; notes generated by bulk extraction tend to be full of them.

A bare #N always selects by position, so an ordinal we hand you in an outline can never be shadowed by a heading that happens to be titled #2. Such a heading stays reachable via the path form (Parent/#2) or via its own ordinal.

The outline is itself bounded by the cap: a note with thousands of headings gets a truncated listing that reports how many sections were omitted (omitted) and the full ordinal range (first_ordinal, last_ordinal), rather than an outline larger than the content window it accompanies. Metadata that does not fit its budget is dropped whole and reported in metadata_omissions — never cut short and never marked inside the field itself, so nothing in a note-controlled field is ever a prefix or server prose. frontmatter_yaml is the frontmatter block's YAML source with the fence lines removed, LF-normalized (the same declared terminator residual content carries); it is the authoritative copy, and the frontmatter JSON view beside it is a convenience that is omitted, with a reason, when YAML holds something JSON cannot say.

limit can lower the cap for a single call but never raise it. If your clients genuinely want larger reads, raise MAX_READ_RESPONSE_CHARS — that is an operator decision, made once, by someone who knows the deployment.

Upgrading: three visible contract changes.

read_note on a large note used to return the whole thing; it now truncates. The response is self-describing, so an agent needs no prior knowledge to continue, but a script that assumed whole-note reads should either pass section= or raise the cap.

And read_note used to return one rendered string — a # <title> / **Path:** header, a \n---\n separator, then the content. It now returns fields, because every component of that header was note-controlled: a note could forge the separator, so an agent recovering the section body by splitting the response could recover a crafted string and write it back over the section. A client that parsed the old envelope must read content (and, for section reads, heading) instead; clients that ignore structuredContent still get an unambiguous JSON text block.

Panel sessions are now server-side rows, so everyone is signed out once at that upgrade. A cookie issued before it carries no session identifier, and such a cookie is refused rather than grandfathered — accepting it would keep the old replay window open for another seven days after the fix shipped. Sign in again; there is nothing to migrate.

Rate limits

The consumer of this server is an agent, and a retry-storming or prompt-injected agent is an ordinary input. Three controls bound how fast one credential can create work.

• A general bucket — MCP_RATE_LIMIT_PER_MINUTE (600) sustained, MCP_RATE_LIMIT_BURST (120) capacity — on every tool call.
• A write bucket — 120/min, burst 30 — that the eight vault-mutating tools must pass in addition, and that PUT /transfer/upload consumes too, charged to the principal that minted the capability so the write rate cannot be escaped by minting links and redeeming them.
• A per-address budget on failed /mcp authentication — 60 failures per 5 minutes — checked before the credential lookup, so a refused probe costs no database query.

The bucket is per principal: an API key, or an OAuth grant. Refreshing an access token continues the same allowance rather than minting a fresh one, and two separate /authorize approvals for the same client hold independent allowances.

What an agent actually sees. A refusal is an ordinary tool result — never a protocol error, never a silent empty result set — and it ends with one machine-readable line:

Error: this credential exceeded its general rate limit of 600 calls per minute, so the call was refused before it ran. Nothing was read, written, or counted against the daily quota. Retry in 1 seconds, or slow the calling loop down.MCP-REFUSAL {"code":"rate_limited","scope":"principal","limit":600,"limit_unit":"calls_per_minute","retry_after_seconds":1}

The MCP-REFUSAL sentinel is line-initial and the JSON is one line, so it survives being quoted into a transcript. A structured tool returns the identical text in its declared error field. retry_after_seconds is present only where waiting can actually help — a refusal for an unassigned vault or an unencodable argument omits it rather than invite a loop that cannot end. The same shape covers the daily quota (over_quota), the query length cap (argument_too_long), and tool-body refusals such as not_found, already_exists, and invalid_path. A partial write also carries a typed outcome: read its explanation before retrying, because some bytes may already have changed. Empty search results and successful no-op calls remain successes.

The transport refusals are outside that contract, because there is no tool call to answer: an over-budget unauthenticated request or an enforced MCP request/authentication concurrency refusal gets an HTTP 429 with Retry-After, and so does an over-rate PUT /transfer/upload — which releases its claim rather than consuming it, so the same link is still redeemable once the bucket refills.

Operational notes.

• Limiter state is in-process and is not persisted, so a restart begins with every bucket full. That is sound only because the container runs --workers 1; raising the worker count multiplies every rate above by the worker count.
• Refusals appear on /admin/performance as refusal counts, not in the latency percentiles. Repeated rate and enforced slot refusals are coalesced — one row per credential/tool/scope per MCP_REFUSAL_LOG_INTERVAL_SECONDS, each standing for 1 + suppressed refusals — so that a refusal loop cannot make writing the log the load.
• The velocity defaults are estimates against a small sample. Read /admin/performance for a week before treating any as settled, and disable one by setting it empty, null or none (zero is refused at startup).
• The daily quota is the durable ceiling and it is separate: keys created from now on get DEFAULT_DAILY_REQUEST_LIMIT (5,000) unless they name a limit — a blank panel field gets the default too — and only an administrator can make a key unlimited, explicitly. Keys that already existed keep whatever they had, and OAuth grants have no daily ceiling at all — velocity bounds only.
• Every new key is a new principal with full bursts, so key creation has its own budget: 10 keys per account and 20 per client address per hour, shared by the JSON API and the panel form, and a non-admin account may hold at most 25 active keys.

The rationale lives in docs/architecture/rate-limits.md.

Concurrency admission

Concurrency admission ships with MCP_CONCURRENCY_MODE=shadow. It records pressure under concurrency_shadow on existing usage rows and emits bounded security events for request/authentication pressure. Calls keep their actual outcome, quota accounting and duration. Shadow mode observes current occupancy with zero wait; it does not predict how traffic would behave under enforcement.

In enforce mode, the server limits full MCP requests (including open SSE streams), authentication database sessions, tools, and usage-log writers. Tools pass velocity, vault and argument checks before acquiring slots; daily quota is checked afterward. A rejected tool receives slot_timeout without spending daily quota. Zero wait means immediate admission or refusal; a positive wait uses a bounded queue and one deadline. A retry hint is not a promise that a running call will finish by that time.

The four tool classes each default to one concurrent call: semantic_search uses embedding, find_related uses vector, the eight vault-mutating tools use write, and the remaining tools use other. Global, tenant and principal ceilings default to 4, 3 and 2. OAuth refresh keeps the same principal. Full request and per-bearer ceilings default to 32 and 4, authentication to 2, and usage writers to 1. All settings and queue limits are listed in .env.example.

Startup validates the pool budget as auth + 2 × tools + writers + 4 ≤ 15. The four connections of headroom are shared with panel, OAuth, indexing and transfer work; this arithmetic cannot guarantee availability when those other consumers exhaust it. Shadow mode does not enforce that budget. The controller is in-process and requires the existing single-worker deployment.

Review pressure observations and long-lived stream occupancy before enabling enforce. Choose off to disable concurrency admission; the existing velocity limits and daily quotas still apply. Shadow requires a zero tool wait and never adds a writer wait or drops a usage row because of its observed pressure.

Architecture

┌──────────────┐                       ┌──────────────────────┐│ MCP clients  │   HTTP + Bearer key   │   FastAPI app        ││  Claude Desk │ ────────────────────▶ │  ┌────────────────┐  ││  Claude Code │                       │  │  MCP server    │  ││  n8n agents  │                       │  │  (28 tools)    │  ││  OpenWebUI   │                       │  └─────┬──────────┘  │└──────────────┘                       │        ▼             │                                       │  ┌────────────────┐  │                                       │  │  Services:     │  │                                       │  │  - vault       │  │                                       │  │  - search      │  │                                       │  │  - embeddings  │  │                                       │  │  - links       │  │                                       │  │  - indexer     │  │                                       │  └─────┬──────────┘  │                                       │        ▼             │                                       │  ┌────────────────┐  │                                       │  │ Postgres +     │  │                                       │  │ pgvector       │  │                                       │  └────────────────┘  │                                       └──────────┬───────────┘                                                  ▼                                       ┌────────────────────┐                                       │  Embedding         │                                       │  provider          │                                       │  (Ollama / OpenAI) │                                       └────────────────────┘

Indexing pipeline

.md files in vault    ↓ skip dot-dirsparse frontmatter, extract tags (YAML + inline #hashtags)    ↓ SHA-256 hashskip if unchanged    ↓UPSERT notes_metadata (path, title, tags[], frontmatter JSONB,                       content_hash, tsvector, modified_at)    ↓extract wikilinks/embeds/markdown-links → resolve targets →note_links (source_id, target_id or NULL for dangling)    ↓chunk content (512 tokens, no overlap) → embed via provider →note_embeddings (note_id, chunk_index, chunk_text, embedding[N])    ↓set embedded_content_hash = content_hash

The indexer runs on startup and every INDEX_INTERVAL_SECONDS (5 minutes by default). Hashes are content-only, so the change detector ignores mtime jitter. Stale embeddings are caught by the embedded_content_hash != content_hash mismatch.

Database schema

GIN indexes on content_tsvector and tags[]. B-tree indexes on the hot foreign keys. pgvector HNSW expression index (embedding::halfvec(N)) halfvec_cosine_ops (m=16, ef_construction=64), built when the dimension is ≤ 2000; results are re-ranked by the full-precision distance. Queries set hnsw.ef_search=80 and dedupe per note in Python after a 5x overfetch.

Project layout

src/  main.py             FastAPI app, lifespan, MCP mount  config.py           pydantic-settings  database.py         async SQLAlchemy engine/session  models/db.py        ORM models  mcp_server/         MCP server, tools, auth middleware  services/           vault ops, anchored filesystem, search, FTS,                      embeddings, links, indexer, transfer  transfer/           public /transfer/* capability-redemption routes  auth/               login, sessions, per-request identity context  api/                control-panel REST endpoints  control_panel/      Jinja2 templates and static assets  oauth/              OAuth 2.0 authorization-code flowalembic/              database migrationsscripts/              one-off ops scripts (e.g. reset_embeddings.py)tests/                pytest suite + smoke-test docsopenspec/             change proposals (spec-driven workflow)

Development

pip install -r requirements-dev.txtpytest

The unit-test suite covers the embedding-provider abstraction, OpenAI batching and retry behavior, config validation, and the dimension-mismatch startup check. Network-bound tests use respx to mock httpx, so no real network access is required.

Dependencies are exact pins (==) in requirements*.txt, bumped by Renovate; renovate.json holds mcp below 2.x (the 2.x migration is issue #365, not a routine bump) and digest-pins the Docker base image.

To run the server outside Docker:

DATABASE_URL=... SECRET_KEY=... VAULT_PATH=... uvicorn src.main:app --reload --no-proxy-headers

Make targets

make init             First-time setup (data dirs, .env)make build            Build Docker image (no cache)make build-cached     Build Docker image (with cache)make push             Push the image to the configured registrymake image            Build and pushmake deploy           Build, scan, push, backup, migrate, recreate containermake up / down / restart / shell   Container lifecyclemake logs             Tail container logsmake db-init          Create database, user, and pgvector extensionmake db-migrate       Run alembic migrationsmake db-check         alembic check — schema vs. ORM models (must be clean)make test-schema      Schema gate: migrations vs. models on a throwaway pgvector containermake db-backup        Dump database to backups dir (derived index tables schema-only; BACKUP_FULL=1 dumps their rows too)make db-restore FILE=<path>   Restore from a backupmake reindex          Explain how to trigger a reindex (panel only; there is no headless trigger)make reset-embeddings Drop and recreate embedding column at configured dimmake rebuild-tsvectors Recompute keyword index for FTS_CONFIGS (no embeddings, no API calls)make status           Show container and health statusmake audit            Audit Python dependencies (pip-audit)make trivy            Scan the local image for HIGH/CRITICAL CVEs (SCAN_IMAGE=obsidian-mcp:local for the bundled stacks)make clean            Remove containers and images (data preserved)

make deploy runs the whole pipeline: build, image scan, push, database backup, alembic upgrade head, then recreate the container. Run make test-schema before any deploy that carries a migration, and make db-check after one.

Security notes

• API keys use the omcp_ prefix and are stored as SHA-256 hashes. The raw key is shown exactly once at creation.
• The control panel is intended to sit behind an external auth gateway. The included docker-compose.yml uses Traefik with an OAuth chain. Don't expose /admin directly to the internet.
• Panel sessions are server-side rows. The signed cookie carries a 256-bit random id; the database stores only its SHA-256, so a database dump contains no usable session. Logging out revokes that row, and a password change, an admin reset, a deactivation or a delete revokes every session of the account.
• The OAuth consent screen identifies the client it is asking about: the redirect host the authorization code would be sent to (taken from the URI's hostname, never its netloc, and shown in punycode rather than decoded), the server-generated client id, and the registration date. Every render says the application registered itself and is not verified by this server; a host outside OAUTH_KNOWN_REDIRECT_HOSTS is called out as unrecognised.
• The OpenAI key is rendered on the settings page as key[:8] + "..." + key[-4:] and never appears in full in HTML or JS sources.
• Path traversal is blocked at the service layer, and containment is proved by the kernel: every directory below the vault root is opened with one openat2(RESOLVE_BENEATH | RESOLVE_NO_SYMLINKS | RESOLVE_NO_MAGICLINKS) from an open root descriptor, and the rest of the operation acts on that descriptor rather than re-walking a name.
• Mutating tools act on the path as named. A final component that is a symlink is refused (naming the link's target) instead of being followed, so an in-vault alias cannot redirect a write. Reads still follow links, which is what an alias is for.
• Every path guard also refuses hidden components, so .obsidian, .git, .trash and friends are out of reach of every tool.
• Transfer links carry their token in the URL fragment, which browsers never send, and are redeemed only from an Authorization: Bearer header. Keep header logging off at your reverse proxy and APM. Unknown, expired, consumed and revoked tokens all get one identical 404 from the public routes; precise status comes from the authenticated check_upload tool.
• import_from_url fetches only genuinely public addresses, under an explicit deny list re-applied at every redirect.
• Failed /mcp authentication is budgeted per client address, counted before the credential lookup so a refused probe costs no database session and no query. The address comes from the proxy headers the app trusts, never from a header read directly, and a request with no resolvable address is charged to a shared slot rather than exempted. What it bounds is the database work an unauthenticated caller can force; it is not a defence against guessing a 256-bit key. See Rate limits.
• Parameterized queries everywhere. No string interpolation into SQL.
• Response headers include HSTS, X-Content-Type-Options: nosniff, X-Frame-Options: DENY and Referrer-Policy: no-referrer. The panel, login and consent pages add a per-response nonce Content-Security-Policy with no inline script (PANEL_CSP).
• The app's own hops are checked at startup: the database follows DATABASE_SSL_MODE, and the embedding endpoint must be https or loopback unless EMBEDDING_ALLOW_PLAINTEXT says otherwise. Each start logs one transport line per hop and an internal_transport_plaintext security event for each hop still in cleartext.

Status

Single-author, in active use as the maintainer's personal exocortex (2,500+ notes, multiple connecting agents). Public for anyone who wants to fork it. Issues and PRs welcome but expect opinionated review. This is a working system, not a generic platform.

Listed in the official MCP Registry as io.github.maxkuminov/obsidian-mcp, published from server.json on every release tag, and on Glama. Both listings carry the tagline at the top of this README: a memory system for your AI agents, stored as plain markdown you can open in Obsidian.

License

MIT. See LICENSE.