lockdocs

Exact-version library docs from your lockfile — local, offline, no rate limits.

Reads your lockfile. Answers from the docs and type declarations of the exact version you installed.
npm · PyPI · crates.io · Go. MCP server + CLI in one Rust binary. No account, no API key. MIT.

[](https://github.com/SylphxAI/repomap#agent-readiness-score)

Docs · Quickstart · Tools · Benchmarks · Compare · How it works

_{A real terminal: one question, two projects. The zod 3 project gets .strict(), the zod 4 project gets z.strictObject(), each cited to package@version file:line.}

Quickstart

npx -y @sylphx/lockdocs setup     # add lockdocs to Claude Code, Codex, Cursor, VS Code, Claude Desktop, Windsurf, Gemini CLI

That's it. setup detects the clients you have, writes their MCP config, and prints every change. Run it again and nothing changes. Then ask your agent something like "Use lockdocs: how do I reject unknown keys with the zod we use?"

From a terminal, inside any project:

npx -y @sylphx/lockdocs zod "reject unknown keys"          # docs for the zod version in your lockfile
npx -y @sylphx/lockdocs api axum::Router::route            # exact signature + doc comment
npx -y @sylphx/lockdocs resolve                            # every pinned version, and whether its docs are here
npx -y @sylphx/lockdocs fetch                              # once: add upstream docs at each version's git tag

{
  "mcpServers": {
    "lockdocs": { "command": "npx", "args": ["-y", "@sylphx/lockdocs", "mcp"] }
  }
}

[mcp_servers.lockdocs]
command = "npx"
args = ["-y", "@sylphx/lockdocs", "mcp"]

Why

Your agent writes code against the library version it remembers, not the one you installed. So it calls .dict() on a pydantic 2 model, cookies() without await on Next.js 15, and /:id routes on axum 0.8. Hosted doc servers help, but they guess the version from the prompt, cap free usage, and cannot see your private packages.

lockdocs takes the version question off the table:

• Exact version, zero config. It reads package-lock.json, pnpm-lock.yaml, yarn.lock, bun.lock, Cargo.lock, uv.lock, poetry.lock, Pipfile.lock, requirements*.txt and go.mod. No library IDs, no "use v14" in the prompt.
• Docs that ship with the code. READMEs, changelogs and docs/ folders, plus the API reference in the package itself: .d.ts declarations with JSDoc, Python docstrings and stubs, rustdoc comments, Go doc comments. If it is installed, it is documented, including your private and internal packages.
• Offline and unlimited. Everything is read from node_modules, your virtualenv, ~/.cargo/registry and the Go module cache. Offline after a one-time model download (129 MB, kept as 32 MB); keyword-only mode (LOCKDOCS_EMBED=0) needs no network at all. No account, no rate limit, and nothing about your dependencies leaves your machine.
• Upstream docs at the exact tag, when you want them. Packages like Next.js, Django and FastAPI ship no docs. lockdocs fetch pulls their docs folders from GitHub at the git tag of your pinned version, once, then stays offline.
• Meaning, not just words. Hybrid retrieval: BM25 fused with a small local embedding model (downloaded once, 32 MB on disk), plus API redirects from deprecation notes ("use model_validate instead").
• Small, cited answers. Packed into a token budget (1,200 by default), every section cited as package@version path:line.

What your agent gets

Three tools, cheap enough to call before every unfamiliar API:

Example, in a pydantic 2 project:

$ lockdocs api BaseModel.model_dump --tokens 300
pydantic@2.9.2 · pypi · .venv/lib/python3.12/site-packages · pinned in uv.lock

### pydantic.main.BaseModel.model_dump (method) — pydantic@2.9.2 pydantic/main.py:352
def model_dump(self, *, mode: Literal['json', 'python'] | str = 'python', include: IncEx | None = None, …) -> dict[str, Any]
Usage docs: https://docs.pydantic.dev/2.9/concepts/serialization/#modelmodel_dump

Generate a dictionary representation of the model, optionally specifying which fields to include or exclude.

The same call in a pydantic 1 project answers that pydantic 1.10.18 has no model_dump, and shows the closest documentation instead.

Ecosystems

Legacy copies bundled inside a package (zod/v3 inside zod 4, pydantic/v1 inside pydantic 2) rank below the current API.

Upstream docs and missing packages (opt-in)

lockdocs fetch adds, once, each direct dependency's upstream docs: it finds the GitHub repository in the package's own metadata and the git tag of your pinned version, and downloads only the docs folders at that tag (Markdown, MDX, reStructuredText, docs examples). Answers then cite next@15.1.0 upstream:docs/01-app/.../cookies.mdx:12. See Upstream docs and fetching.

Not installed? Fetch the exact version (opt-in)

Out of the box lockdocs reads only your disk (plus the one-time embedding model download). If a pinned package is not installed (a fresh clone, CI, a lockfile you are reviewing), it says so and tells you how to install it. Pass --fetch (or set LOCKDOCS_FETCH=1, or lockdocs setup --fetch) to let it download exactly that version from the registry (npm tarball, PyPI wheel or sdist, crates.io .crate, Go module proxy zip) into its cache. Fetched answers say fetched from registry.npmjs.org. You can also ask for a version you do not use: lockdocs npm:zod@4.1.5 "strict object" --fetch.

Benchmarks

105 questions whose correct answer depends on the version, over 15 libraries (zod, Next.js, React Router, pydantic, axum, tokio, Tailwind CSS, ESLint, Prisma, React, Vite, Express, SQLAlchemy, Django, FastAPI), each asked in a real project with that version installed. An answer passes when it contains the version-correct API and none of the other version's. Same questions and grader against Context7's anonymous API, on a GitHub-hosted runner (run):

• Where versions matter most, lockdocs wins by 2x (38/45 vs 19/45 on older majors). Context7 often answers with the newest API (pydantic 1 questions get pydantic 2 answers).
• Tied on the newest majors (53/55 each) and on tokio (5/5 each); ahead on held-out questions (16/18 vs 15/18), which were written before the ranking changes they measure and not used for tuning.
• ~27x faster, fewer tokens, no quota. lockdocs latency is a fresh CLI process per question; lockdocs fetch is a one-time download per project.

Method, questions, per-question results and scripts: benchmark page and bench/.

How it compares

Context7 knows many sites and guides that never ship inside a package, so for a question whose answer lives only on a docs website, it can do better. lockdocs is for the other case: when the version matters, when you are offline or rate-limited, and when the package is yours. They combine well.

How it works

1. Resolve. Parse every lockfile in the project (and monorepo roots above it) into exact (ecosystem, name, version) triples, marking direct dependencies.
2. Locate. Find each package's files on disk and check the installed version against the lockfile. Drift is reported, never hidden.
3. Extract. Split READMEs, changelogs and doc folders into heading-scoped sections; parse .d.ts, Python, Rust and Go sources with tree-sitter into symbols with signatures, doc comments and qualified paths (z.object, tokio::task::spawn, pydantic.main.BaseModel.model_dump).
4. Index and cache. BM25 (identifier-aware tokenizer, light stemming, a small programming synonym table) plus one embedding per entry from a static model2vec model, cached on disk by package, version and source, so each version is indexed once, ever.
5. Answer. Fuse keyword and embedding scores, apply docs signals and deprecation redirects, then pack results into the token budget with citations.

Typical costs on a GitHub-hosted runner: indexing every declaration in Next.js 15 (7,700 symbols) takes about 0.4 s, zod 4 about 0.05 s, once per version; a query then takes 5-20 ms.

CLI

lockdocs <package> [question]   Docs for the version this project pins (no question: overview)
lockdocs resolve [filter]       Pinned versions and where their docs are
lockdocs docs <question>        Search all direct dependencies (--pkg to focus)
lockdocs api <symbol>           Exact signature + doc comment
lockdocs fetch [package...]     Once: upstream docs at each version's tag, missing packages, the model
lockdocs index [package]        Build indexes ahead of time
lockdocs cache [clean]          Show or delete the cache
lockdocs setup                  Configure MCP clients (--client a,b --dry-run --remove --fetch)
lockdocs mcp                    MCP server on stdio

Options: -C/--root <dir>, --pkg <package>, --tokens <n>, --fetch, --offline, --json

Prebuilt binaries for macOS (arm64, x64), Linux glibc (x64, arm64) and Windows x64 ship through npm; each GitHub release has them too. From source: cargo install --git https://github.com/SylphxAI/lockdocs lockdocs.

Privacy

lockdocs reads files on your machine and answers over stdio. Network use: the embedding model once from huggingface.co (pinned revision, SHA-256 checked; LOCKDOCS_EMBED=0 or --offline skips it), and, only when you run lockdocs fetch or enable fetching, public registries and GitHub for the exact package versions requested. Nothing about your project is sent. The cache lives in your OS cache directory (LOCKDOCS_CACHE overrides it).

Also from Sylphx

• repomap: a map of your codebase for AI agents: code graph, search, call paths and change impact, with an interactive graph UI.
• anymd: any file to clean Markdown for your AI agent: PDF, Word, PowerPoint, Excel, EPUB, HTML, images, audio and video.
• readme-mark: beautiful README images from one URL.

Star history

License

MIT © Sylphx