36 lines
1.2 KiB
Markdown
36 lines
1.2 KiB
Markdown
# inference-docs-mcp
|
|
|
|
The Inference Cooperative documentation as an **MCP server** — members opt in
|
|
per conversation and can search the co-op docs (getting started, models, API
|
|
access, charter, policies) from the chat.
|
|
|
|
## Design
|
|
|
|
- **MCP over stdio**, implemented with Python stdlib only (newline-delimited
|
|
JSON-RPC 2.0) — no dependencies to install.
|
|
- The corpus is fetched at runtime from the public `co-op/docs` repo
|
|
(Gitea raw API), chunked by `##` heading, and ranked with BM25. The index
|
|
rebuilds lazily when older than 30 minutes, so doc edits show up without
|
|
any redeploy.
|
|
- Runs as a child process of the chat frontend (LibreChat spawns it via its
|
|
`mcpServers` config with `type: stdio`).
|
|
|
|
## Deploy (LibreChat sandbox)
|
|
|
|
```bash
|
|
cloudron push --app <chat-app-id> server.py /app/data/inference-docs-mcp/server.py
|
|
# librechat.yaml:
|
|
# mcpServers:
|
|
# inference-docs:
|
|
# type: stdio
|
|
# command: python3
|
|
# args: ["/app/data/inference-docs-mcp/server.py"]
|
|
```
|
|
|
|
Restart the chat app; the tool (`search_docs`) registers on boot.
|
|
|
|
## Tool
|
|
|
|
`search_docs(query, limit?)` → the most relevant doc sections with their
|
|
doc title + heading context.
|