trovex

[trovex: one canonical doc for your coding agents, ~60% fewer tokens.]

trovex

trovex: one canonical doc for your coding agents, ~60% fewer tokens.

[Join the tsukumo community on Discord]

Your coding agents (Claude Code, Cursor, Windsurf, Zed, any MCP client) reread the repo every session to work out which .md is current, then answer from a guess. You pay for that on every session, every agent, every teammate.

trovex indexes your repo's markdown and exposes one MCP tool. Your agent asks a question; trovex returns the single current doc that answers it as a path:line pointer with a freshness marker (canonical / stale / duplicate), and serves just the section that answers instead of the whole file. Agents also write what they learn back through one shared point, so every agent and teammate reads the same source of truth instead of re-deriving it.

About 60% fewer tokens on doc lookups, measured at equal task-success on our own repo (it varies by yours). Runs locally: vectors in SQLite, embeddings via ONNX, no cloud or API keys.

What you're installing

A 30-second trust check, since the decision happens on the README, not the directory listing:

• First-party, open source. Built and run in production by tsukumo, the team behind it. AGPL-3.0-or-later.
• Local-first, nothing leaves your machine. Vectors in SQLite, embeddings via ONNX. No cloud, no API keys, no network call to answer a query.
• Confined writes, no shell. Six MCP tools: three read-only (trovex, trovex_read, trovex_search) and three that mutate only trovex's own doc store (trovex_write, trovex_tag, trovex_delete). No shell execution, no writes to your source files.
• The ~60% is reproducible. Measured at equal task-success on our own repo (median 69%, ~41–81% by repo, n=26, LLM-judged); trovex search prints the savings on yours. Full method at trovex.dev/measure.

Quick start

trovex is in public beta, on PyPI. No clone needed; uv tool install puts trovex on your PATH:

uv tool install trovex   # one-time, no clonetrovex setup             # wire into Claude Code (skill + hooks + MCP); idempotent
trovex index /path/to/your/repo        # index your markdown (~1 min)trovex search "how do we roll back a deploy?"   # ask, prints the tokens it savedtrovex serve                           # wire into your agent: MCP at /mcp, dashboard at /savings

Don't have uv? It's a one-line install: curl -LsSf https://astral.sh/uv/install.sh | sh (or brew install uv).

The search step is the fast way to see the point: it returns the one canonical doc and prints how many tokens that saved versus reading the top few candidates. Once trovex is wired into your agent over MCP, the same numbers accumulate on the savings dashboard at http://localhost:8765/savings.

Prefer not to install anything yet? uvx trovex search "..." runs a single command in a throwaway environment, no install.

Wire it into your agent

trovex is an MCP server. Point your client at http://localhost:8765/mcp after trovex serve. Per-client setup (Claude Code, Cursor, Windsurf, Cline, Zed, Roo) is at trovex.dev/for.

Claude Code, one command. trovex setup installs the Claude Code skill, the Active-Memory hooks, and registers the MCP server in one step (idempotent, safe to re-run):

trovex setup

Restart Claude Code afterwards so it loads the skill + hooks. Prefer to wire just the MCP server by hand? trovex setup --no-skill --no-hooks, or:

claude mcp add --transport http trovex http://localhost:8765/mcp

For Cursor, one click (after trovex serve): Add trovex to Cursor

How it works

trovex turns your repo's markdown into one queryable, canonical store, then serves each agent the single current doc that answers a question, not a pile of candidates to rank.

flowchart LR  A["Your repo<br/>.md files"] -->|trovex index| B["Chunk + parse"]  B --> C["Embed locally<br/>ONNX · bge-small"]  C --> D[("sqlite-vec<br/>vector store")]  Q["Agent question"] -->|MCP| E["Route + rerank"]  D --> E  E --> F["One canonical doc<br/>path:line + freshness"]  F --> G["Agent answers from<br/>the current doc"]

Four ideas do the work:

• Canonical, not complete. For each question there should be one doc that answers it. trovex marks every doc canonical / stale / duplicate, so retrieval can prefer the one that's still true: a stale doc is exactly as similar to a query as the current one, so similarity alone can't tell them apart.
• Route, then serve the section. A query returns a path:line pointer to the one doc and just the section that answers, not the whole file, and not the top-k pile your agent would otherwise read and rank itself. Closing that gap is where the tokens are saved.
• Local by default. Indexing, embeddings (ONNX, bge-small-en-v1.5) and vector search (sqlite-vec) all run on your machine. No cloud, no API key, no network call to answer a query.
• A shared write-back path. Agents store what they learn once (trovex_write) in trovex's own store; every other agent and teammate reads it back (trovex_read) instead of re-deriving it.

The read-and-write loop that keeps every agent on the same source of truth:

sequenceDiagram  participant A as Agent  participant T as trovex (MCP)  participant S as Store (sqlite-vec)  A->>T: trovex("how do we roll back a deploy?")  T->>S: route + rerank candidates  S-->>T: one canonical doc  T-->>A: path:line + section + freshness  A->>T: trovex_write("rollback runbook: …")  T->>S: upsert (dedupes near-copies)  Note over S: the next agent reads it back via trovex_read

Prove it on your own repo

The ~60% is a claim you can run, not a number to take on faith. Two commands:

trovex bench /path/to/your/repo          # token-accounting model, instant, no LLM, no keytrovex bench /path/to/your/repo --eval   # full answer+judge A/B at equal task-success (needs OPENAI_API_KEY)

bench reports the distribution (median + spread), not a best case: the cost of reading the one routed canonical doc versus the top-k candidates an unaided agent would read. --eval goes further: both arms answer, an LLM judges, and a saving counts only when both answer correctly. Full method and our own numbers are at trovex.dev/measure.

Already running trovex through the md-guard hook? trovex measure compares your real .md token consumption before and after, from the hook's baseline log.

Dev notes: the claims eval harness

bench --eval above answers "does trovex still get the right answer" per-query, one arm vs the other. trovex eval-harness (src/trovex/eval_harness.py) is the release-gate version of that idea: a versioned cases.jsonl set, retrieval quality (hit@k/MRR/recall@k, free) plus a weighted rubric score (correctness35/autonomy25/actionability20/safety10/concision10, blind — the judge never sees which config produced an answer), a $ budget cap, and resumability.

make eval                                                          # retrieval-only, no key, CI-safeuv run trovex eval-harness . --gate --budget-usd 1.0 --resume run.jsonl   # full rubric pass, needs OPENAI_API_KEY

--retrieval-only (what make eval runs) never calls an LLM — it gates purely on hit@1 against benchmarks/token-savings/eval-baseline.json. The full rubric pass needs a key, so it's a manual/CI-secret-gated run, not part of make test. The bundled cases.jsonl (47 cases) and eval-baseline.json thresholds are versioned in benchmarks/token-savings/ — the baseline is a placeholder until it's been run once for real and the numbers copied in deliberately (the file is never auto-overwritten by a run).

MCP tools

• trovex(q): route a question to the right on-disk .md and get back path:line pointers with freshness markers, not a pile of files to rank.
• trovex_write(content, kind?, doc_id?, tags?, section?) / trovex_read(query | doc_id, section?): docs owned inside trovex. An agent stores a record (an incident, a decision, "what actually worked") once; every other agent and a second dev read it back as content (optionally just one section) instead of re-deriving it. Pass section= to trovex_write to patch one heading's section in place instead of replacing the whole doc.
• trovex_search(query, k?, tags?): passage-level retrieval across the store with tag filters, for when you want the top matching chunks rather than one canonical doc.
• trovex_tag(...) / trovex_delete(...): tag or soft-delete a stored doc; delete is a recoverable archive, not a hard wipe. Both touch only the trovex store, never your files.

Humans read trovex-owned docs at /doc/{id} in the rendered reader. To make agents route .md writes through trovex_write instead of the disk, install the PreToolUse hook deploy/hooks/trovex-md-guard.sh and carve out exceptions in .trovexignore.

How it compares

• vs CLAUDE.md / AGENTS.md: one static file that goes stale and can't route a question to the right doc, vs many docs kept canonical and served per query. More.
• vs repomix / files-to-prompt: pack the whole repo into the window vs retrieve the one answer. More.
• vs a vector DB / plain RAG: a ranked pile of candidate chunks with no freshness signal vs one current doc with an explicit marker. More.

The reasoning behind the ~60% number is written up in the benchmark methodology.

Stack

• Python 3.11 + uv
• FastAPI (MCP HTTP + server-rendered HTML UI)
• fastembed (local embeddings, ONNX under the hood)
• sqlite-vec (vector search in SQLite)
• Jinja2 + HTMX (UI, no build step)

Embeddings: local by default, bring your own

trovex embeds locally out of the box with BAAI/bge-small-en-v1.5 (ONNX, 384-d). No API key, nothing leaves your machine. You can swap in any embedder:

• Another local model: TROVEX_EMBED_MODEL=<fastembed model> plus TROVEX_EMBED_DIM=<its dimension> if it's not a built-in.
• An OpenAI-compatible endpoint (incl. a local server like Ollama, LM Studio, vLLM): TROVEX_EMBED_PROVIDER=openai, TROVEX_EMBED_MODEL=<model>, TROVEX_OPENAI_BASE_URL=http://localhost:11434/v1, TROVEX_EMBED_DIM=<dim>. Point it at localhost and you stay fully local; point it at OpenAI (text-embedding-3-large) for stronger retrieval at the cost of sending each chunk to OpenAI's API.

Changing the model changes the vector dimension, so switching requires a reindex.

Public beta

trovex is in public beta. Install it, run it on your repo, and if it saves you tokens a GitHub star helps other devs find it. Issues and PRs welcome.

Questions, or comparing notes with other people running agents? Join the community on Discord: tsukumo.ch/discord.

Security

trovex is local-first and single-tenant: it runs on your machine, indexes your docs, and serves your agents. Mutations are gated behind the X-TROVEX-Write-Token header, and the default is fail-closed: with no token configured, trovex auto-generates a per-instance token on first run and persists it to <data_dir>/.write_token (chmod 600), so a network-exposed instance does not accept anonymous writes. Set TROVEX_WRITE_TOKEN to share one token across machines, or TROVEX_ALLOW_UNAUTH_WRITES=1 to deliberately run with open writes on a trusted localhost. The trust model, what's hardened, and how to report a vulnerability are documented in SECURITY.md.

License

trovex is licensed under the GNU AGPL-3.0-or-later (see LICENSE). You can self-host and modify it freely; if you run a modified version as a network service, AGPL requires you to share your changes.

Part of the suite

trovex is the context layer of a four-part open-source suite for running AI coding agents in production, built by tsukumo:

• trovex, context: serve agents the one canonical doc per query instead of rereading the repo (~60% fewer tokens per lookup; method at trovex.dev/measure).
• wrai.th, orchestration: run and coordinate a fleet of agents.
• yoru, observability: session receipts of what each agent actually did.
• dokan, deterministic execution: run the agent's settled, repeatable work as scripts in clean containers, no model in the loop.

Working with a team?

trovex is free to run yourself. If your team is rolling out coding agents at scale and wants hands-on help doing it well, or to embed and host a modified trovex privately without the AGPL's copyleft obligations, that's what the consulting is for. Reach out to tsukumo, the team behind trovex.