tldr: openground is a tool that lets you give controlled access to documentation to AI agents. Everything happens on-device.

I'm sharing openground, an opensource and completely on-device RAG tool that let's you give controlled documentation access to your coding agents. Solutions like Context7 MCP provide a likely source of truth for docs, but their closed-source ingestion and querying pose security/privacy risks. openground aims to give users full control over what content is available to agents and how it is ingested. Find a documentation source (git repo or sitemap), add it to openground via the CLI, and openground will use a local embedding model and vector db (lancedb) to store your docs. You can then use the CLI to install the MCP server to your agent to allow the agent to query the docs via hybrid BM25 full-text and vector search.

Here's an example of giving Claude code access to a specific version of FastAPI docs:

uv tool install openground openground add fastapi --source https://github.com/fastapi/fastapi/tree/master/docs/en --version 0.126.0 openground install-mcp --claude-code // or --cursor or --opencode

This is still an early release, but I've implemented several useful features so far:

- docs ingestion from git repos (only .md and other text files, no code) or sitemap.xml - specific version ingestion for git sources - easy `install-mcp --<agent-name>` command for popular agents like opencode, claude code, and cursor - efficient(ish) updating of docs that have already been saved

This is still an early version, so expect breaking changes. Things I am working on:

- command that keeps docs updated on a schedule - project specific access control from the MCP server - docs "registry" to allow pushing and pulling of documentation embeddings to and from S3 - lighter-weight package - better docs

Suggestions and PRs welcome! I'll also be around for discussion.