Anthropic Docs

Offline mirror + MCP server for the entire Anthropic documentation universe — Claude API, Claude Code, Model Context Protocol, and the Claude Help Center. Search 1,100+ pages locally from Claude Code, Claude Desktop, or any MCP client, with git-based tracking of what the official docs changed each week.

Why: LLMs' built-in knowledge of Anthropic's own APIs is chronically outdated, and the docs move fast (weekly changelogs, new APIs, deprecations). A local, refreshable mirror gives your agent authoritative answers — and git diff tells you exactly what Anthropic changed, including edits that never make it into a changelog.

No HTML scraping involved: everything comes from the official LLM-friendly endpoints (llms.txt indexes + per-page .md).

Quick start

Requires Python 3.9+ (crawler, stdlib-only) and uv (server; resolves its own dependencies).

git clone https://github.com/YUYUqueen/anthropic-docs-mcp.git
cd anthropic-docs-mcp

# 1. Build the corpus (~1,100 pages, ~50 MB, a few minutes)
python3 crawl.py

# 2. Register the MCP server with Claude Code (available in all projects)
claude mcp add --scope user anthropic-docs -- uv run "$(pwd)/mcp_server.py"

For Claude Desktop, add to your MCP config:

{
  "mcpServers": {
    "anthropic-docs": {
      "command": "uv",
      "args": ["run", "/absolute/path/to/anthropic-docs-mcp/mcp_server.py"]
    }
  }
}

Tools

Tool	What it does
search_docs	Full-text search with title-weighted ranking, snippets, per-site filtering, pagination — plus semantic layer/doc_type filters and native-language matching when metadata exists
read_doc	Read one page by relative path; long docs paginate automatically
find_related	Look up every page about an entity (e.g. "prompt caching"): primary pages, related pages, co-occurring entities (requires metadata)
recent_changes	What the official docs changed recently, from the corpus's git history

Optional: semantic metadata layer

If you have the Claude Code CLI installed, you can enrich the corpus with per-page structured metadata — extraction runs through your Claude subscription (headless claude -p with Haiku), so there is no API bill:

python3 extract_metadata.py --limit 8   # sample first
python3 extract_metadata.py            # full corpus (~1,100 calls, incremental, resumable)
python3 build_topics.py                # aggregate into TOPICS.md + entity_index.json

This unlocks: layer/doc_type filters on search_docs, one-line summaries in results (language set via SUMMARY_LANG, default English), native-language queries matching English docs, the find_related reverse index, and a content-first TOPICS.md topic map. Re-run both scripts after each crawl; only changed pages are re-extracted.

What gets crawled

Directory	Source	Scope
docs/platform/	platform.claude.com	All guides + API reference for REST, Python, and TypeScript (7 other SDK languages excluded — edit SDK_LANGS_EXCLUDED in crawl.py to change)
docs/claude-code/	code.claude.com/docs	Everything
docs/mcp/	modelcontextprotocol.io	Everything
docs/support/	support.claude.com	English articles

English only by design: it is the only complete, always-current locale, and your MCP client's model can answer in any language anyway.

Keeping it fresh & tracking changes

The crawler is incremental (content-hash comparison; unchanged pages are skipped) and safe to re-run. To also get recent_changes and a reviewable history of what Anthropic edits each week, keep the corpus in its own directory with its own git repo (in this repo docs/ is gitignored so Anthropic's content never ends up in forks):

export ANTHROPIC_DOCS_DIR=~/anthropic-docs   # both scripts honor this
git init "$ANTHROPIC_DOCS_DIR"               # once
python3 crawl.py --git-commit                # crawl + commit a snapshot when anything changed

Schedule it weekly (cron, launchd, or a Claude Code scheduled task). Then, inside the corpus directory:

git log --oneline               # every docs snapshot
git diff HEAD~1 --stat -- docs/ # what changed since last crawl
git diff HEAD~1 -- indexes/     # pages Anthropic added or removed

ANTHROPIC_DOCS_DIR overrides the corpus location for both scripts (defaults to the repo directory).

Copyright note

This repository contains tooling only. The crawled content belongs to Anthropic / the Model Context Protocol project — build your mirror locally and don't redistribute the docs/ output (it's .gitignored for that reason).

License

MIT — see LICENSE. Not affiliated with Anthropic.