mcp-agentdocs
# agentdocs
**Fresh, source-verified docs for building MCP servers and Claude agents — served straight to your AI coding agent over MCP.**
[](https://www.npmjs.com/package/mcp-agentdocs)
[](./LICENSE)
[](https://modelcontextprotocol.io)
Coding agents hallucinate on fast-moving APIs like the Model Context Protocol and the Claude Agent SDK: training data lags the spec, so you get plausible-but-wrong imports, deprecated patterns, and invented options. `agentdocs` is a local MCP server that gives your agent a curated, **source-verified** corpus to search instead of guessing.
- 🔎 Three tools: `list_topics`, `search_docs`, `get_doc`
- 📎 Every answer is attributed to its upstream source URL
- 🏠 Runs **locally** over stdio — no account, no hosting, no data leaves your machine
- ✅ Each doc carries a `verified` flag set only after human review against the source
## Install & connect
Requires Node.js 18+.
### Claude Code
```bash
claude mcp add agentdocs -- npx -y mcp-agentdocs@latest
```
### Cursor / other MCP clients (`mcp.json`)
```json
{
"mcpServers": {
"agentdocs": {
"command": "npx",
"args": ["-y", "mcp-agentdocs@latest"]
}
}
}
```
Then ask your agent things like *"check agentdocs for how MCP stdio handles logging"* or
*"search agentdocs for the Streamable HTTP session header"* — it will call the tools and
answer from the verified corpus.
## Tools
| Tool | Use it to |
|------|-----------|
| `list_topics` | See every curated doc, its topic, source URL, and verified status |
| `search_docs(query, topic?)` | Get ranked, source-attributed snippets for a question |
| `get_doc(id)` | Read the full verified markdown of one document |
## What's covered (v1)
The Model Context Protocol (spec 2025-11-25) and building MCP servers / Claude agents.
Corpus grows over time; run `list_topics` for the current set.
## Freshness
Every document records the date it was last verified against its primary source. Sources
are diffed on a regular cadence; when one drifts, the affected doc is re-verified and a new
version is published. Using `mcp-agentdocs@latest` (as above) keeps you on the current corpus.
See [CHANGELOG.md](./CHANGELOG.md) for what changed and when.
## Development
```bash
npm install
npm run build # tsc -> dist/
npm test # vitest: retrieval, manifest integrity, unicode safety
npm run ingest # refresh sources into corpus/_staging/ for human review
```
Curation is the point: `npm run ingest` fetches upstream sources for review but never
silently overwrites a curated doc — a human diffs, promotes, and sets `verified: true`
in `corpus/manifest.json` only after checking the content against its source. Run
`npm run check-freshness` to flag docs that have drifted or are due for re-verification.
## License
MIT
TDQS
Scored across 3 tools
Each tool has a clearly distinct purpose: list_topics lists all available documents, search_docs searches for specific content, and get_doc retrieves full document. No overlap or ambiguity.
All tool names follow a consistent verb_noun snake_case pattern: list_topics, search_docs, get_doc. Easy to predict and understand.
With only 3 tools, the set is minimal but covers the essential operations for a documentation server: browse, search, and retrieve. It is slightly under the typical range but still reasonable for a focused server.
The tool surface covers the full lifecycle of accessing curated documentation: listing all documents, searching for specific content, and retrieving full documents. No obvious gaps for the stated purpose.