Plenty of tools can index module source. What is unusual here is indexing consumer repositories too, so the answer to "how do we tag things" comes from what your teams actually wrote rather than from the module's documentation. Two things to settle before adopting: the licence is BUSL 1.1, so production use is a commercial conversation, and authentication is disabled by default — turn it on before the instance is reachable by anyone else.
A knowledge base over your Terraform module repositories. It clones them, parses every HCL file, embeds the results into PostgreSQL with pgvector, and then indexes consumer repositories to learn how the modules are really used — naming, variables, tagging, layout. The MCP endpoint puts that base in front of an agent.
- Nine tools — `query_modules`, `pick_modules`, `list_modules`, `get_module_details`, `get_dependencies`, `get_module_usage`, `find_similar_usages`, `fetch_example_code`, `get_stats`
- `list_modules` also takes a `semantic_query` for natural-language search rather than exact names
- `get_dependencies` walks the dependency tree using PostgreSQL recursive CTEs, so you can trace what depends on what
- `get_module_usage` and `find_similar_usages` answer 'how do we normally call this' from real consumer code
- A knowledge layer distils conventions across six dimensions — naming, variables, tagging, layout, versions and deployment — and treats them as authoritative in prompts
- Version tracking through automatic git tag discovery, with code-hash caching so re-indexing does not repeat LLM and embedding calls
- A full audit trail of LLM, MCP and API calls
You host it. `git clone`, `cp .env.example .env`, set `POSTGRES_PASSWORD`, `JWT_SECRET` and an LLM API key, then `docker compose up -d` — the frontend on port 3000, the API on port 8000, the MCP endpoint at `http://localhost:8000/mcp/`. Index a repository by POSTing its `repo_url` and branch to `/index/`; private repos need an SSH deploy key at `./worker_deploy_key` or a path in `SSH_KEY_PATH`. The LLM backend is your choice — Anthropic directly or through Bedrock, OpenRouter, Ollama, or any OpenAI-compatible endpoint, selected with `LLM_BASE_URL` and `LLM_MODEL`. Authentication is off by default, with local email/password plus JWT or ALB-terminated SSO as alternatives. Licensed under BUSL 1.1: non-production use is permitted, production use needs a commercial licence until it converts to AGPL-3.0 in 2029.
One command plus a key — npx -y mcp-remote https://terraform-rag.io/mcp, then supply credentials