jDocMunch MCP
jDocMunch is an MCP server for coding agents that retrieves the exact documentation section a task needs, without loading whole files into the context window.
Index a documentation set once by heading hierarchy, then fetch a single section, a heading subtree, or a ranked search result — extracted byte-precisely from the original file.
Install · Quickstart · Benchmarks · Commercial licensing
[PyPI version]
[PyPI - Python Version]
[License]
[MCP]
[Local-first]
[DOI]
Free for personal use. Commercial use requires a paid license — terms below.
Why jDocMunch?
The problem. An agent asked "how do I configure authentication?" opens a documentation file, skims hundreds of paragraphs it does not need, opens another, and repeats. Large context windows do not fix this. They just make the waste affordable enough to ignore until the bill arrives, and they crowd out the context the model actually needed.
The mechanism. jDocMunch parses a documentation set into a section tree keyed by heading hierarchy, stores each section's byte offsets into the original file, and exposes retrieval over MCP. Sections keep durable identities across re-indexing as long as path, heading text, and heading level are unchanged.
The outcome. The unit of access changes from file to section. An agent retrieves the installation section, one configuration block, or a specific heading subtree — and nothing else.
What makes it different
Section-first retrieval
Search and retrieve documentation by section, not just file path or keyword match.
Byte-precise extraction
Full content is pulled on demand from exact byte offsets into the original file.
Stable section IDs
Sections retain durable identities across re-indexing when path, heading text, and heading level remain unchanged.
Evidence
Four benchmarks against public documentation corpora, each with the corpus, date, and per-query results recorded in benchmarks/.
Read these as per-corpus results, not as a single headline multiple. Savings depend on how large the containing file is relative to the section you needed: a small file with one heading saves almost nothing, and the Kubernetes corpus saves a great deal. The benchmark files record the queries that did poorly alongside the ones that did well.
A separate, measured result from the v1.121.0 projection work, on this repository's own docs at max_results=10: a search row went 1,989 chars → 319 with compact=true (−84%), or 431 with snippet_bytes=200 (−78%) while removing the follow-up get_section call entirely.
Retrieval quality is gated, not assumed. Every release runs a replay fixture over a frozen golden set and fails below nDCG 0.95. That gate has failed builds and blocked releases; it is not decorative.
Install
Requirements: Python 3.10+, any MCP-compatible client.
No virtualenv to manage, nothing written into system Python, and it works as-is on PEP 668 distros (Ubuntu 24.04+, Debian 12+) where bare pip install is refused. Don't have uv yet?
init detects your MCP clients, writes their config entries, installs the doc-exploration prompt policy so your agent actually reaches for the tools, and optionally installs hooks and indexes your docs.
Other install paths
Verify:
Manual Claude Code setup:
No install step — uvx fetches and runs the server on demand. Prefer it on your PATH (and required for hooks)? uv tool install jdocmunch-mcp, then claude mcp add -s user jdocmunch jdocmunch-mcp.
Installing the server makes the tools available; it does not break an agent's habit of brute-reading files. One line in your CLAUDE.md does that:
Quickstart
Assumes: jDocMunch installed and registered with your client, and a folder of documentation.
Index a local documentation folder:
It prints JSON naming the corpus and what it found:
section_count greater than file_count is the whole point: the index addresses headings, not files.
Then, inside your agent:
Using jdocmunch, search the docs for "authentication configuration" and show me that section.
The agent should call search_sections, then get_section on the top hit — returning one section rather than a file. _meta.tokens_saved on the get_section response is a modeled saving: the containing document's size minus the section served (_meta.savings_basis: "modeled"). search_sections itself earns no credit.
Next step: get_toc_tree for a structural view of the whole corpus, or index_repo to index documentation straight from a GitHub repository.
Very large folders
An MCP client stops waiting on a tool call after a fixed time. Indexing a folder with tens of thousands of files can take longer than that. On one Windows machine, a full index of 10,000 files (271,585 sections) took 67 s without embeddings, and sentence-transformers embeddings added about 3 minutes.
If a folder is that large, or an index_local call times out, index it from a shell:
The CLI runs the same indexing with no tool timeout. It writes to the same store as the MCP server, so the server uses the index once the command finishes. If you set DOC_INDEX_PATH in your client's MCP config, set it in the shell too.
- Custom names. If the folder was indexed under a custom name, pass the same
--name. Without it, the CLI uses the folder name and writes a second index.
- The file cap. A folder is capped at 10,000 files, newest first. Raise it with
--max-files N (the CLI form of the MCP tool's max_files). When a call hits the cap, its warning prints the command with the folder's discovered count.
- Narrowing a folder. Exclude what you don't need with
--extra-ignore-pattern (repeatable, the CLI form of extra_ignore_patterns), or index only listed files with --paths-from FILE (the CLI form of paths).
- Embeddings.
--no-embeddings builds a lexical index first. Run the command again later with --rebuild to add embeddings.
Scanned PDFs
A scanned PDF is a picture of text, with no text layer to index. The optional [ocr] extra reads them with an OCR engine already on your machine. jDocMunch never installs or downloads an engine, a language pack or a model.
Engines. jDocMunch looks for one in this order and uses the first that answers:
JDOCMUNCH_TESSERACT_PATH, if you set it.tesseract on PATH.- Tesseract's default install folders:
C:\Program Files\Tesseract-OCR, the per-user folder under %LOCALAPPDATA%\Programs, /opt/homebrew/bin, /usr/local/bin and /usr/bin.
- On Windows, the built-in Windows OCR, if an English OCR language is installed (it is on English-language Windows).
JDOCMUNCH_OCR_ENGINE=tesseract or windows picks one. JDOCMUNCH_OCR_LANG passes Tesseract languages that are already installed (default eng).
OCR needs two keys. It runs only when both are set:
- You set
JDOCMUNCH_OCR=1 in your MCP client's env block (or your shell, for the CLI).
- The call asks for it:
index_local(ocr=true), or index-local --ocr on the CLI.
So an agent can't turn OCR on by itself. The choice is stored with the index, and later refreshes and the watcher reuse it; ocr=false or --no-ocr turns it off. Every response that ran with OCR requested has an ocr block saying whether it was enabled in config, whether it was used, how much it read and with which engine.
What you get. Each scanned image becomes a ## Page N section. Every section from OCR is marked text_source: "ocr", with ocr_engine, ocr_pages and (Tesseract only) ocr_confidence. The document opens with a visible note that its text came from OCR, and the response warns about it. Any part of a PDF with its own text layer keeps it. OCR sections earn no token-savings credit.
How good it is. On 6 real public-domain scans of printed prose (Encyclopaedia Britannica 1911 and The American Printer), Tesseract 5.4 found 97.4% of the words with a 2.0% character error rate, and Windows OCR 97.1% and 2.0%. The limits:
- Tables and forms may come out scrambled. The words are mostly there, so a search can find them, but columns and fields lose their layout. On two scanned forms, Tesseract found 25% to 80% of the words.
- Decorative and black-letter (Gothic) type isn't read. Neither engine reads it.
- Handwriting isn't read.
Speed. Tesseract took 3.4 s per image here (Windows OCR 0.3 s): a 20-image scan took 69 s with Tesseract and 6 s with Windows OCR. An MCP call gives OCR a 60-second budget (ocr_budget_seconds), so with Tesseract, a scan of 20 images or more may not finish in one call. It's skipped as ocr_budget_exhausted. What was already read is cached, and the next call carries on. For a folder of scans, run the CLI, which has no budget:
The first 50 scanned images of a document are read (ocr_max_pages, --ocr-max-pages); the rest are reported as ocr_pages_capped.
Privacy. OCR runs on your machine. Tesseract is a local program, and Windows OCR is an on-device Windows API; no page leaves the machine. But OCR turns scans into searchable text. If you OCR tax forms, pay stubs, IDs or medical records, their contents (names, SSNs, account numbers) become readable to any agent that can query this index, and they're stored in the index's text cache on disk. jDocMunch's secret-file rules match file names, not contents, so they won't catch a scanned W-2. Keep such files out of indexed folders, or exclude them with extra_ignore_patterns.
What you can do
- Retrieve one section instead of a document.
get_section and get_sections pull byte-precise content from the original file; get_section_excerpt narrows further.
- Search by meaning, not just keywords.
search_sections fuses BM25 with semantic cosine when an embedding provider is configured. compact=true, fields=[...], and snippet_bytes=N cut the response further.
- Navigate structure.
get_toc, get_toc_tree, get_section_path, get_section_descendants, and section_neighbors traverse the heading tree without reading content.
- Find what documentation is missing or rotting.
get_doc_coverage, get_undocumented_symbols, get_stale_pages, get_orphan_sections, get_broken_links, and doc_health_radar.
- Work across API specs.
find_endpoint, list_endpoints_by_tag, find_operations_using_schema, and get_schema_graph treat OpenAPI documents as first-class.
- Preflight documentation changes.
check_section_delete_safe and get_section_blast_radius before you remove or restructure.
- Know when an answer is stale. Content reads disclose
_meta.freshness, _meta.verdict, and which source layer answered.
64 tools in total. The full reference is in USER_GUIDE.md.
How it works
Everything runs locally. Indexes live under your home directory; no hosted service is required for indexing or retrieval.
- Parsing is per format, one module each: Markdown/MDX, reStructuredText, AsciiDoc, Org-mode, Jupyter notebooks, HTML, plain text, OpenAPI (YAML), JSON/JSONC, XML/SVG/XHTML, Godot scenes, and — via the optional
[office] extra — PDF, DOCX, PPTX, and EPUB.
- Storage is a versioned local index (
INDEX_VERSION = 3) that auto-migrates on first load. A 1.x release never forces a reindex.
- Retrieval is lexical BM25 by default, hybrid when embeddings are available.
- Embeddings are optional and provider-agnostic — Gemini, OpenAI, an OpenAI-compatible endpoint, or a local offline model through either FastEmbed (ONNX) or sentence-transformers (torch). Without one, search stays lexical and entirely offline.
Deeper detail: ARCHITECTURE.md and SPEC.md.
Memory in long-running processes
A parsed index takes about 2.5 to 3 times its file's size in memory (measured on five real indexes, 0.5 MB to 112 MB). Two settings keep long-running processes bounded:
JDOCMUNCH_INDEX_CACHE_MAX_BYTES caps the indexes the server keeps parsed between calls. The default is 512 MiB, estimated as 3 times each index file's size. Past it, the least recently used index is dropped and re-read when it's next needed. A single index larger than the cap is still kept. 0 or less turns the byte cap off; the 8-index cap stays either way.JDOCMUNCH_WATCH_INPROCESS=1 makes the watcher re-index in its own process instead of a child per batch (see below).
Security and privacy
Local-first by design. Your documentation is parsed and stored on your machine, and the base package's only default network behavior is an anonymous savings counter — a random ID plus aggregate token counts, no content, no paths, no PII.
Opt out completely:
Embedding and summarizer providers call their configured API only when you enable them, and never by default. watch-install registers a login service only when you run it yourself.
Background behavior, fully disclosed
A model download, the first time a local embedding provider runs. Both
offline providers fetch their model from HuggingFace on first use and cache it
on disk. Nothing downloads until you enable embeddings, and a lexical-only
install never contacts the hub. Startup warmup is skipped when the model is
not already cached, so a first run defers the download to your first search
rather than stalling the MCP handshake behind it
(#110).
A package install and a model download, only if you say yes to init.
jdocmunch-mcp init checks whether an embedding provider is available. If none
is, it asks once whether to turn semantic search on. The default answer is
no. On a yes it does two things and prints both before doing them:
- runs
python -m pip install "fastembed>=0.8.0" in the Python environment
jdocmunch is running in (about 160 MB installed; pulls onnxruntime, not torch);
- downloads one model file,
all-MiniLM-L6-v2 in ONNX form, from
huggingface.co (87 MB, once).
After that, embedding runs on your machine. No text from your documents is sent
anywhere. --yes does not answer this question for you: it accepts config
edits, not a download. A scripted install opts in with --with-embeddings.
--dry-run prints the command and the sizes and touches nothing. If pip is not
usable (a uv tool or pipx install has none) or the install fails, init
prints the manual command for your install style and carries on with word
matching. The MCP server itself never installs or downloads anything unasked.
Why init asks at all: on a public benchmark of 492 real user questions over 6
documentation sets, semantic (hybrid) search put a direct answer in the top 5
results for 47% of questions, against 34% for word matching alone. Method, data
and the negative result that came with it:
jdoc-rerank-bench. Measured on
one CPU-only machine, embedding added about 7 seconds per 1,000 sections to a
first index. doc_list_repos reports has_embeddings per index, so you can see
which ones are still words-only.
FastEmbed as the offline provider. pip install jdocmunch-mcp[fastembed]
runs the same all-MiniLM-L6-v2 model through onnxruntime instead of torch,
which is a much smaller install. When both offline providers are present
FastEmbed is preferred; JDOCMUNCH_EMBEDDING_PROVIDER=sentence-transformers
selects the other one. On the shared model the two write the same vector
store, so switching runtimes does not re-embed your corpus. Point FastEmbed at
a different model with JDOCMUNCH_FASTEMBED_MODEL and it keeps its own vectors
instead, because vectors from two models are not interchangeable
(#126).
Where the FastEmbed model lives, and repairing it. FastEmbed's own default
folder is <temp dir>/fastembed_cache, and OS cleanup can empty it. When the
model can't load, index_local says so in embedding_provider_error and a
warning, and search is word matching until it is fixed. init checks the model
from local files only, without a download. If the model is missing or damaged,
it offers a repair, and the repair asks for the same consent as the install: a
yes at the prompt, or --with-embeddings. The repair downloads the model
(87 MB, once) into ~/.cache/jdocmunch-mcp/fastembed, outside the temp
directory. The server loads from that folder once the model is there; an
install whose temp-dir copy works keeps using it and is never re-downloaded.
JDOCMUNCH_FASTEMBED_CACHE names a different folder, and FastEmbed's own
FASTEMBED_CACHE_PATH still takes precedence over both.
A local OCR program, only if you turn OCR on. With the [ocr] extra,
JDOCMUNCH_OCR=1 in config and ocr=true on the call (or --ocr),
jDocMunch runs Tesseract as a short-lived child process once per scanned page,
or calls Windows' on-device OCR. Neither opens a network connection. The text
is cached under your doc-index directory (.office_cache/ocr). Nothing runs in
the background and nothing is installed. See Scanned PDFs for
what OCR'd text exposes.
A child process per watcher batch. The watcher (jdocmunch-mcp watch, or
the jdocmunch-watch login service) re-indexes each batch of changed files in a
short-lived child, python -m jdocmunch_mcp.watch_batch. A re-index rewrites
the whole index file, and memory used for that doesn't return to the system in
a long-running process, so the watcher's memory used to climb toward its limit
(#189). The child
returns it when it exits. It:
- runs once per batch, and exits when the batch is done;
- opens no network connection of its own; it calls only what indexing already
calls, such as your embedding or summary provider if you set one;
- starts its embedding provider fresh each time. On one Windows PC, a
one-file batch took 0.3 s in-process, 0.55 s in a child, and 1.7 s in a child
with FastEmbed on (the model loads each time). sentence-transformers wasn't
measured and loads more slowly;
- on Windows, opens no console window.
JDOCMUNCH_WATCH_INPROCESS=1 restores the previous in-process re-index.
A child process, when local embeddings are in use. When the
sentence-transformers provider is active, jDocMunch runs the embedding model
in a child process (python -m jdocmunch_mcp.embeddings.worker) instead of
inside the server. It:
- starts when something first needs an embedding — at startup if the model is
already in your local HuggingFace cache, otherwise on the first search or
index that uses it. A lexical-only install never spawns it;
- opens no network connection and speaks only to its parent, over a private pipe;
- exits when the server exits, and is killed if it stops responding;
- is not a login service, is not registered anywhere, and survives nothing.
This exists because importing the embedding stack inside the server process can
deadlock in the Windows loader
(#118), hanging every
tool call for as long as the server runs. Disable it with
JDOCMUNCH_EMBED_WORKER=0, which restores the previous in-process import.
A login service, only if you install one. jdocmunch-mcp watch-install
registers the doc watcher to start at login (systemd user unit, launchd agent,
or a Task Scheduler task named jdocmunch-watch). Nothing installs it for you.
Once installed it:
- re-indexes every locally-indexed doc repo when a doc file on disk changes;
- runs exactly
jdocmunch-mcp watch with the flags you passed to
watch-install — --no-ai-summaries to keep the summarizer out of it,
--quiet to suppress its per-change log lines;
- writes to
watch.log and watch.err under your doc-index directory;
- is removed by
jdocmunch-mcp watch-uninstall.
⚠ Re-running watch-install rewrites the service definition, so a
hand-edited one is replaced. It now prints what it replaced; pass the flags to
watch-install itself so an upgrade keeps them
(#120).
Path traversal prevention, symlink escape protection, secret exclusion, file-size limits, binary detection, and encoding safety are documented in SECURITY.md, along with how to report a vulnerability.
Limitations
- Section retrieval helps least on small files. If a document has one heading and 40 lines, retrieving the section and reading the file cost about the same.
- Semantic search requires an embedding provider. Without one, search is lexical only — good for identifiers and exact phrasing, weaker for paraphrased questions.
- Office formats need the optional
[office] extra and are supported for local indexing only. A scanned (image-only) PDF has no text layer, so without OCR it is not indexed. The response names it under skipped_paths.no_extractable_text, adds a warning, and reports coverage_complete: false.
- OCR is optional, off by default, and modest. With the
[ocr] extra, an engine already on your machine and both keys set, scanned pages are read (see Scanned PDFs). Printed prose reads well. Tables and forms may come out scrambled, and decorative, black-letter and handwritten text isn't read.
- Freshness is disclosed, not guaranteed. A section whose source cannot be checked is reported as
unknown rather than assumed current.
- jDocMunch does not parse code. Symbols, signatures, and call graphs belong to jcodemunch-mcp; tabular data belongs to jdatamunch-mcp.
Documentation
Licensing and commercial use
Released under the jDocMunch-MCP Dual-Use License (full terms). Free for non-commercial use. Commercial use requires a paid license, one-time, sold by jMunch LLC.
jDocMunch only: Builder, $29 (1 developer) · Studio, $99 (up to 5) · Platform, $499 (org-wide internal deployment)
Full jMunch suite (code + docs + data): Trio Builder, $99 · Trio Studio, $449 · Trio Platform, $2,499
Personal and other non-commercial use is free. Work use at a for-profit company, or for client or paid work, needs a license, paid once.
1.x compatibility commitment
Every 1.x license entitles you to every future 1.x release. We will never ship a 1.x version that:
- removes or renames an MCP tool (deprecated tool names keep their aliases),
- drops a
Section field from the response shape,
- forces a reindex without auto-migrating your existing index on first load,
- changes the JSON wire format of any tool response in a way that breaks an existing consumer,
- or makes a previously-default behavior raise.
Anything that would require breaking these promises is reserved for a future major version (2.x). The full machine-checked contract is enforced via tests/test_server.py (tool-name and required-field invariants) and the replay-fixture gate that runs on every release.
Support and project status
Actively maintained. Issues and bug reports: GitHub Issues. Security reports: see SECURITY.md. Commercial licensing questions go through jcodemunch.com.
Part of the jMunch suite alongside jcodemunch-mcp (code symbols) and jdatamunch-mcp (tabular data). All three implement jMRI, the open retrieval interface spec.
