OpenDocuments MCP
A self-hosted RAG platform: cited answers over GitHub, Notion, Drive and files, with an MCP mode for Claude Code.
- 123 ★ on the repository
- Last commit
Extended access
Why this level
- Through MCP the agent gets access to the whole indexed document corpus.
- With a cloud model, document fragments go to the provider.
Install
- Works in
Manual install
npm install -g opendocumentsAfter install run `opendocuments init` and `opendocuments start`; for MCP mode run `opendocuments start --mcp-only`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Through MCP the agent gets access to the whole indexed document corpus.
- With a cloud model, document fragments go to the provider.
This is third-party code. Review the repository files before installing.
What it does
OpenDocuments turns scattered documents into an AI-searchable knowledge base on your own infrastructure. Connectors pull GitHub, Notion, Google Drive, Confluence, S3 and GCS, Swagger and OpenAPI, web pages and local files. Parsers read Markdown, PDF, DOCX, XLSX, CSV, code and other formats, metadata goes to SQLite, vectors to LanceDB, and answers arrive with source citations. The `opendocuments start --mcp-only` command turns on MCP mode so Claude Code, Cursor or Windsurf search the document corpus while writing code. Models are local through Ollama or cloud ones: OpenAI, Anthropic, Google and xAI. There are a CLI, web UI, HTTP API, SDK, plugins and a team mode with keys and roles.
Who it is for. Teams whose knowledge sits in GitHub, Notion, Confluence and drives and who want to give an agent search over it without sending data to a third-party cloud.
Good fit when
- An agent writing code needs internal authentication and API documents.
- Onboarding guides in DOCX on S3 and specs in Notion are searched with one question and citations.
- Sensitive documents stay inside: a local model through Ollama answers without a cloud.
- Several departments share one instance with workspace isolation and an audit log.
Not a fit when
- You need search over a single folder without deploying a platform: a lighter local MCP server fits better.
- No model is configured: the web UI blocks questions until the corpus and model are ready.
- Documents cannot go to cloud models and there is no local model on the machine.
Example request
Find in our documents how the JWT token refresh works and show the sources.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
OPENDOCUMENTS_MODEL_PROVIDER- Model provider, for example `ollama`.
OPENDOCUMENTS_MODEL_BASE_URL- Model server address, for example `http://localhost:11434`.
Limitations
Node.js 20 or newer is required. A model is needed: local through Ollama or cloud with a key; with a cloud model, document fragments go to the provider. A plain `start` never downloads a model. With an unprepared corpus the web UI blocks questions and shows the fix step. In `--mcp-only` mode the client gets access to the whole indexed corpus, so index only what the agent may see. The last repository commit was in July 2026.
How to disable. Remove the `opendocuments` entry from the MCP client configuration and run `npm uninstall -g opendocuments`. Delete the SQLite and LanceDB index in the project directory separately.
FAQ
How do I connect it to Claude Code?
Add an MCP server with `command` set to `opendocuments` and `args` set to `["start", "--mcp-only"]`. The agent then searches the indexed documents.
Which sources are supported?
GitHub, Notion, Google Drive, Confluence, S3 and GCS, Swagger and OpenAPI, web pages, uploads and local files. `opendocuments connector sync github` triggers a connector sync.
Can it run fully locally?
Yes. With Ollama, a local model and embeddings answer without a cloud; the `opendocuments init` wizard detects Ollama and offers to pull missing models.
What do I check if there are no answers?
Run `opendocuments doctor`: it checks Ollama connectivity, model availability and configuration validity.
How do I back it up?
The backup command snapshots SQLite and LanceDB, and an instance recovers with one command.
Discussion
No questions or reviews yet.
Related
MCP serversA local document base for your agent: hybrid semantic and keyword search plus a web UI for browsing and uploads
OpenWiki builds linked Markdown repository wikis, retains source evidence and lets agents search documentation across projects.
The MCP server connects an agent to a learning canvas where a question about a passage opens a linked document.
An MCP server built on Docling: converts PDFs into structured documents, builds new ones and feeds them to RAG
A local MCP server lets an agent search your Zotero library, Obsidian vault and Apple Notes through a C++ search core
An MCP server on top of pandoc: the agent converts Markdown, HTML, DOCX, LaTeX, EPUB, ODT and other formats into one another
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| OpenDocuments (this card) | no | not required | 123 | |
| MCP Local RAG | no | not required | 429 | mcp-local-rag searches a local folder without an external database; OpenDocuments adds GitHub, Notion and Confluence connectors, a web UI and a team mode |
| knowledge-rag | no | not required | 301 | knowledge-rag turns a folder into a knowledge base for Claude Code; OpenDocuments holds a corpus from many external sources with roles and audit |
README in short
The platform follows the standard RAG cycle: source, parsing, chunking, embeddings, reranked retrieval and a cited answer. MCP mode exposes the same search to agents, and the rest is available through the CLI, web UI and HTTP API.