kabi-rag
v0.2.9
Published
Per-project RAG memory MCP server for AI coding agents (bundled native Go binary)
Maintainers
Readme
Kabi RAG
Kabi RAG gives AI coding agents per-project RAG memory. The package includes a native server and browser dashboard; npm is only the installer and launcher. You do not need Go to install it.
npm install -g kabi-rag
kabi-rag versionStart the dashboard and setup wizard:
kabi ragkabi-rag --serve also works. In an interactive terminal, kabi-rag without
arguments opens the dashboard server; when launched by an MCP client over pipes,
no arguments selects MCP stdio mode.
Open http://localhost:7777. The main setup flow offers all supported providers. Optional Quick Setup detects already running local endpoints; it does not install, start, or stop database/model services. Use the rescan button after starting a service yourself. Detection checks Ollama on 11434; OpenAI-compatible servers on 8082, 8080, 8083, 8090, 1234; TEI on 8081/8080; Qdrant on 6333/6334; and PostgreSQL on 5432/5433. Other ports, remote hosts, and authenticated endpoints can be entered through the main provider flow.
For PostgreSQL Quick Setup, enter the existing database, username, and password.
Kabi RAG validates login and pgvector in that database before saving; it never
creates a database or enables pgvector for you. After saving, stop and restart
Kabi RAG with the same arguments and environment to activate the configuration.
For the default listener, restart with kabi rag; custom ports and RAG_STATE_DIR
must be preserved. Saving alone does not reinitialize the running provider.
You need one vector store (Qdrant, PostgreSQL/pgvector, or Chroma) and one embedding provider (Voyage AI, self-hosted TEI, or an OpenAI-compatible endpoint). PostgreSQL and TEI are alternative components, not requirements to install together. The binary retains the original project's indexing, search, migration, and MCP features.
The npm package bundles binaries for Linux and macOS (x64 and ARM64) and Windows (x64). Installation does not download from the source repository.
The default HTTP listener is loopback. If an admin token is configured (saved or
RAG_ADMIN_TOKEN), open the dashboard and sign in; API/MCP clients use a bearer
token. Dashboard sessions expire after 12 hours and do not survive server restart.
Never disable authentication just to make the dashboard work.
Remote listeners require an admin token unless --allow-insecure-no-auth is
explicitly selected (unsafe). Serve remote traffic over HTTPS using
--tls-cert server.crt --tls-key server.key; proxies must preserve Host and
re-encrypt upstream to the HTTPS listener.
For a separate test installation, use a disposable local npm prefix/cache and an
absolute RAG_STATE_DIR for config and metrics. Also supply a disposable database
and credentials: a different HTTP port alone does not isolate server state.
Release packaging
Release binaries are staged and packed on Linux by CI. Unix binaries are marked executable and checked before publishing. The launcher also repairs lost Unix execute bits in writable installations, including installs with scripts disabled. Build the dashboard before rebuilding and staging all five native binaries; checksum validation alone does not establish source freshness. A published npm version cannot be reused: select a new version and update the changelog before staging the final release artifacts.
