Claude Code Documentation Tool
Official Claude docs, always up-to-date, always at your fingertips. Stop searching the web — ask Claude directly and get accurate answers grounded in official documentation.
Why Use This?
Claude knows a lot — but documentation changes fast. API parameters shift, new features land, SDK methods get renamed. This tool gives Claude a live index of every official doc page and fetches the latest page from Anthropic's servers on demand, so answers come from the source, not stale training data. The repository itself stores only metadata — a page manifest and a lossy search index — never documentation prose.
| Without claude-code-docs | With claude-code-docs |
|---|
| Claude guesses from training data | Claude reads the latest official docs |
| Broken or outdated URLs in answers | Correct platform.claude.com / code.claude.com links |
| "I think the API works like..." | "According to the documentation..." |
| You verify answers manually | Answers cite specific doc pages |
How It Works
The repository commits only metadata — a page manifest and a prose-free search index (a couple of MB, no documentation text). Your machine fetches the actual .md pages from Anthropic's servers on demand into a local cache. So:
- Current — answers come from the live page on
platform.claude.com / code.claude.com, not a snapshot or stale training data.
- Private & compliant — no Anthropic documentation is redistributed. The repo commits none of it (not at the tip, not in history); you fetch it yourself, straight from the source.
- Tiny — the clone is a few MB of metadata, not the ~100 MB doc corpus, and pages are cached only as you open them.
Quick Start — Plugin Install (Recommended)
Two commands, no dependencies:
/plugin marketplace add costiash/claude-code-docs
/plugin install claude-docs@claude-code-docs
That's it. On your next session Claude will automatically:
- Clone a tiny metadata repo to
~/.claude-code-docs/ (manifest + search index, no prose)
- Fetch documentation pages into a local cache in the background, and keep both current each session
- Make the
/docs skill available for manual lookups
- Activate the auto-discovery Skill — Claude reads docs automatically when you ask Claude-related questions
What the Plugin Gets You
/docs skill — Look up any topic: /docs hooks, /docs extended thinking, /docs Agent SDK sessions- Auto-discovery Skill — Claude proactively searches docs when you ask about Claude Code, the API, SDKs, or prompt engineering. No
/docs prefix needed.
- Interactive Courses — Turn any Claude topic into a stunning, self-contained HTML course with animated diagrams, protocol conversations, quizzes, and code translations. Just say "create a course on hooks" or run
/docs --course <topic>.
- Session-start auto-updates — Docs stay fresh automatically. No cron jobs, no manual pulls.
- Content search — Shell-based full-text search and fuzzy filename matching, no Python needed
- Nothing to install — the client is pure
bash + curl + jq (already on macOS/Linux), no Python runtime needed
Legacy: Script Install (Migration)
For environments without plugin support, the install script clones documentation locally:
curl -fsSL https://raw.githubusercontent.com/costiash/claude-code-docs/main/install.sh | bash
If Claude Code is detected, the script will guide you to install the plugin instead. Without Claude Code, it falls back to git clone for basic documentation access.
Team / Organization Adoption
Auto-prompt every team member to install the plugin by adding this to your project's .claude/settings.json:
{
"extraKnownMarketplaces": {
"claude-code-docs": {
"source": {
"source": "github",
"repo": "costiash/claude-code-docs"
}
}
},
"enabledPlugins": {
"claude-docs@claude-code-docs": true
}
}
Commit this file to your repository. When a team member trusts the project folder, they'll be prompted to install the marketplace and plugin automatically — no manual setup needed.
Interactive Courses — Learn Claude by Doing
This isn't just a docs mirror anymore. Ask about any Claude topic and get a beautifully crafted, interactive HTML course — complete with animated protocol diagrams, hands-on quizzes, code-with-English translations, and a stunning dark Obsidian & Amber visual theme. One self-contained file, zero setup, opens right in your browser.
/docs --course hooks # Generate a course on Claude Code hooks
/docs --course tool use # Deep dive into the Tool Use API
/docs --course prompt caching # Master caching with visual explanations
Or just ask naturally after any docs response — Claude will offer to create a course on whatever you just looked up.
https://github.com/user-attachments/assets/e36ae4c1-2ee6-4932-b0a5-3463cd20e012
What you get:
- 4-7 scroll-based modules with progressive learning arc
- Protocol Conversations — animated chat-style visualizations showing how Client, API, Claude, and tools actually exchange messages
- Code translations — real API examples from the docs with line-by-line plain English explanations
- Quizzes that test application — "What would you use for this scenario?" not "Define this term"
- Glossary tooltips on every Claude-specific term
- Obsidian & Amber theme — dark, atmospheric design with Instrument Serif headings, warm amber accents, grain textures, and glass effects. Not your typical tutorial.
Courses are saved to ~/.claude-code-docs/courses/ so you can revisit or share them with your team.
Changelog Reports
Stay on top of what's changing in Claude's documentation. Generate a visual HTML report of recent doc updates — grouped by category, with summaries and key highlights:
/docs --report # Last 7 days of changes
/docs --report 24h # Last 24 hours
/docs --report 30d # Last 30 days
Each entry in the report includes a "Create Course" button. Click it to copy a course generation command to your clipboard — paste it into Claude Code to instantly deep-dive into any topic that caught your eye.
Usage
Direct Lookups
/docs hooks # Claude Code hooks
/docs mcp # MCP server configuration
/docs agent sdk python # Agent SDK Python guide
/docs --course hooks # Generate an interactive course on hooks
/docs --report # HTML changelog of recent doc changes
/docs -t # Check freshness (read-only)
/docs what's new # Recent documentation changes
Natural Language Queries
The /docs skill understands intent — ask questions in plain English:
/docs what are the best practices for Agent SDK in Python?
/docs explain the differences between hooks and MCP
/docs how do I configure extended thinking for the API?
/docs show me all prompt library templates
Claude finds the right docs, reads them, and synthesizes a clear answer with source links.
With the Auto-Discovery Skill (Plugin Only)
When installed as a plugin, you don't even need /docs. Just ask naturally:
"How do I set up MCP servers in Claude Code?"
Claude recognizes this is a documentation question and automatically reads the relevant docs before answering.
Documentation Coverage
The full Claude documentation set — every page on platform.claude.com and code.claude.com — re-indexed every 3 hours:
- API Reference — Messages API, Admin API, multi-language SDKs (Python, TypeScript, Go, Java, Ruby, C#, PHP)
- Agent SDK — Python and TypeScript SDK guides, sessions, hooks, custom tools
- Claude Code — CLI docs: hooks, skills, MCP, plugins, settings, sub-agents
- Agents & Tools — MCP connectors, tool use patterns, agent capabilities
- Core Documentation — Guides, tutorials, prompt engineering, extended thinking
- About Claude — Model capabilities, context windows, pricing
- Getting Started — Quickstart guides
- Testing & Evaluation — Eval frameworks, testing guides
- Release Notes — Version history and changelogs
How Updates Work
- Automatic (Plugin) — Each session the metadata syncs (
git reset --hard origin/main) and a background fetch updates only changed pages in the local cache
- Automatic (CI/CD) — GitHub Actions regenerates the manifest + index from Anthropic's
llms.txt + sitemaps every 3 hours
- On-Demand —
/docs sync fetches changed pages now; /docs -t checks freshness
- Safe — Manifest safeguards prevent catastrophic changes (min 200 pages discovered, max 10% removed per sync, min 250 pages)
Upgrading from a v1 (mirror) install? Nothing to do — your next session auto-migrates. The SessionStart hook hard-syncs to the metadata-only v2 layout (removing the old committed docs/, preserving your cache and courses), then fetches pages on demand.
Troubleshooting
| Problem | Solution |
|---|
/docs not found | Restart Claude Code so the plugin loads; for a script install, check ls ~/.claude-code-docs/ |
| Docs seem outdated | /docs sync (or the SessionStart hook, which hard-syncs each session) forces an update; /docs -t only checks freshness |
| Plugin not working | Run /plugin list to verify installation |
| Script install fails | Install the plugin instead: /plugin install claude-docs@claude-code-docs |
Uninstalling
Plugin:
/plugin uninstall claude-docs@claude-code-docs
Script install:
~/.claude-code-docs/uninstall.sh
Security
- No documentation prose is committed — the repo holds only metadata (manifest + lossy index)
- The client fetch layer enforces a domain allowlist (
code.claude.com, platform.claude.com, raw.githubusercontent.com), HTTPS-only, with atomic writes and sha256 integrity checks
- Manifest safeguards prevent catastrophic changes; input sanitization and path-traversal protection on filenames
- Full test suite (including a mocked-curl harness for the client fetch layer)
Contributing
See CONTRIBUTING.md for architecture overview, development setup, testing requirements, and PR guidelines.
Acknowledgments
License
MIT License. Documentation content belongs to Anthropic. Tool code is open source — see LICENSE.
