Auto-updating archive of Anthropic's builder documentation: every source that publishes markdown, fetched 4x daily. 4,100 files from six live sources, plus a frozen archive of the engineering blog.
Clone this repo and point Claude Code at it. Every doc, tutorial, cookbook, skill, and engineering post Anthropic publishes on a markdown surface -- searchable, version-controlled, and offline.
git clone https://github.com/thevibeworks/claude-code-docs
cd claude-code-docsThen ask Claude Code anything:
claude "how does the agent loop work?"
claude "show me the MCP spec for tool annotations"
claude "what did the 'Building Effective Agents' post recommend?"
claude "how do I set up hooks in the Agent SDK?"Counts are files on disk as of 2026-09-09; uv run scripts/fetcher.py --tree
prints them live.
| Source | Section | Files | What |
|---|---|---|---|
| code.claude.com | --section claude-code |
198 | Claude Code + Agent SDK docs |
| platform.claude.com | --section api |
2,228 | API reference, build guides |
| claude.com/docs | --section products |
226 | Claude Tag, Cowork, office agents, connectors |
| modelcontextprotocol.io | --section mcp |
347 | MCP spec, SDKs, governance |
| github.com/anthropics | --section github |
764 | Cookbooks, skills, plugins, courses, SDK docs |
| support.claude.com | --section support |
372 | Help articles |
| anthropic.com | frozen 2026-07-08 | 158 | Engineering, research, news posts -- see below |
anthropic.com has no --section flag: it is HTML-only, the jina.ai proxy the
fetcher used was removed in July 2026, and content/blog/ has been a static
archive since 2026-07-08. What is in it, against what anthropic.com's sitemap
lists today:
content/blog/ |
Archived | Upstream | Coverage |
|---|---|---|---|
engineering/ |
25 | 25 | complete |
product/ |
4 | 4 | complete |
research/ |
72 | 155 | 46% |
news/ |
57 | 260 | 22% |
The engineering posts -- "Building Effective Agents", context engineering, tool use -- are all here. Research and news are not, and will not grow until something converts HTML to markdown.
content/
en/docs/claude-code/ Claude Code + Agent SDK
en/api/ API reference (1,500+ endpoints)
en/build-with-claude/ Platform features
en/agents-and-tools/ Tool use, agent skills
en/manage-claude/ Admin, billing, managed agents
claude/ Product docs (Claude Tag, Cowork, office agents)
mcp/ MCP protocol spec + community
blog/ frozen 2026-07-08, not refreshed
engineering/ Building Effective Agents, context engineering, ...
research/ Research papers
news/ Model releases
product/ Product announcements
github/
cookbooks/ 164 recipes + notebooks
skills/ 90 official Agent Skills
plugins-official/ 266 plugin docs
courses/ 80 prompt engineering notebooks
quickstarts/ Deployable app starters
code-action/ GitHub Actions for Claude Code
sdk-python/ Python SDK reference
sdk-typescript/ TypeScript SDK reference
support/ 365 help articles
Auto-updates four times daily via GitHub Actions. To fetch manually:
# Requires: uv (https://docs.astral.sh/uv/)
uv run scripts/fetcher.py # Fetch everything (~3 min)
uv run scripts/fetcher.py --section mcp # MCP spec only
uv run scripts/fetcher.py --section github # GitHub repos only
uv run scripts/fetcher.py --incremental # Skip existing files
uv run scripts/fetcher.py --tree # Show all sources + counts
uv run scripts/fetcher.py --discover # Probe domains for new sourcesGitHub repo fetching needs GITHUB_TOKEN or GH_TOKEN in the environment.
Every fetched source serves a .md variant of each page, so nothing is
converted from HTML. That is also why content/blog/ is frozen: anthropic.com
is HTML-only and the jina.ai proxy path it used was removed in July 2026.
See sources.json for the complete machine-readable source
registry.
The fetcher doesn't just download from hardcoded URLs. It probes every known
Anthropic domain for robots.txt, sitemap.xml, llms.txt, and — the
question that decides everything — whether the domain serves .md variants at
all. It also enumerates github.com/anthropics and watches the Location
header on every redirect it follows.
Known domains: anthropic.com, platform.claude.com, code.claude.com,
support.claude.com, modelcontextprotocol.io, claude.ai, claude.com,
academy.claude.com
The result is a file, not a log line. Every full run rewrites
discovery.json with what exists upstream that
sources.json does not. This matters because printing it did not work: the
pipeline had been logging support.claude.com -> academy.claude.com four
times a day for weeks, into an Actions log nobody opens, and those 725 pages
were eventually found by a human chasing a dead support article. A discovery
that isn't a diff doesn't reach anyone.
So a new domain, a new anthropics repo, or a domain that starts serving
markdown now shows up as a tracked change, gets classified as high-signal, and
opens a PR — the same path a new doc takes. discovery.json.review is the
actionable list: reachable, serves markdown, nothing fetches it. Empty is
healthy. Adding a source stays a human decision.
uv run scripts/fetcher.py --discover # manual probe; writes the same fileSitemaps are treated as incomplete, not authoritative. Upstream de-indexes
pages it still serves: in July 2026 platform.claude.com dropped every
per-language SDK reference page from both its sitemap and its llms.txt while
continuing to edit them, and the archive quietly stopped refreshing 1,560 files
for seven weeks. So every full run also refetches what is already on disk, and
pages that really died are removed by the reaper below rather than by absence
from an index.
A page removed upstream used to live here forever — the fetcher only ever added or overwrote. Full runs now delete archived files whose URL returns 404/410 or the site's HTML shell, with two guardrails:
- Only markup is deleted automatically. A file holding real markdown whose URL has died is content Anthropic removed and we may hold the only copy; it is reported for a human instead of destroyed by a job that merges its own PRs.
- A mass-deletion circuit breaker. More than 200 pages vanishing at once means an upstream outage, not 200 real deletions — nothing is deleted and the run fails loudly.
--no-reap reports what would go without touching anything.
tombstones.json records every URL confirmed gone upstream, with the date and
reason. It exists so that a page which died once does not report as a fresh
failure on every subsequent run — 123 standing failures would pin the success
rate at 96.9% and bury the one new breakage that matters. Later runs count
known deaths quietly and print only what changed: pages newly gone, and pages
that came back (whose tombstone is then removed automatically). The success
rate is computed over live docs, so it means something.
Two GitHub Actions workflows power this repo:
fetch-claude-docs.yml -- Scheduled every 6 hours. Runs the fetcher, then hands the diff to Claude Code (via claude-code-action) which decides: ignore noise, commit minor fixes directly, or create a PR for meaningful changes. Sends push notifications via barkme for PRs.
claude-review.yml --
Triggered on PRs and @claude mentions. Reviews changes, merges routine
updates, creates tracking issues for version bumps, and alerts humans only
for breaking changes.
PRs welcome. The fetcher is a single Python file (scripts/fetcher.py)
with no framework dependencies beyond aiohttp and aiofiles.
To add a new source:
- Add the source definition to
sources.json - Add the fetch logic to
scripts/fetcher.py - Run
uv run scripts/fetcher.py --section <name>to test - Update
--sectionchoices in the CLI
Unofficial mirror for educational and development purposes. Documentation content is sourced from Anthropic's public sites. For official docs, visit code.claude.com and platform.claude.com. Repository code from anthropics/claude-code. Redistribution should comply with Anthropic's terms.