lilbee MCP
lilbee runs local models across several GPUs and gives agents MCP search over your files with a file and line citation
- 70 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The program indexes chosen files and web pages, runs local models and starts a local HTTP server
- Models download from Hugging Face, and with a key set requests go to a cloud provider
Install
- Works in
Manual install
brew tap tobocop2/lilbee && brew install lilbeeA self-contained build. Also available: npm install -g lilbee, AUR, Docker, Nix, Flatpak, Scoop and a binary from the releases. Then lilbee self-check.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The program indexes chosen files and web pages, runs local models and starts a local HTTP server
- Models download from Hugging Face, and with a key set requests go to a cloud provider
This is third-party code. Review the repository files before installing.
What it does
lilbee is one program that runs local models (chat, embedding, vision, rerank), places them across every GPU you have and answers questions over your files, notes, code and web pages with an exact file and line citation. It needs no separate model server, vector database or container: lilbee keeps the index itself. It runs on `llama.cpp` with Metal, Vulkan and CUDA, and is reached as a terminal app, CLI, MCP server, HTTP API or Python library. There are two paths for coding agents. `lilbee launch opencode`, `lilbee launch hermes` and `lilbee launch claude` set an agent up on local models in one command: for opencode and hermes lilbee registers itself as a provider and an MCP server, while Claude Code gets an Anthropic-compatible API through environment variables with its settings untouched. The second path: the agent calls `lilbee_search` and gets cited snippets from your library, so it does not invent function names. The system decides placement itself: a large chat model is spread across the fewest cards that fit, with headroom on each. Files and index stay on the machine, there is no telemetry or account, and a cloud provider is called only with your key. The project is in active beta and every PyPI release is a pre-release.
Who it is for. People who run Claude Code, OpenCode or Hermes on local models and want to give the agent search over their own docs, dependencies and code with no cloud.
Good fit when
- You want to run Claude Code on a local model with one `lilbee launch claude` command
- The agent invents function names and needs docs and dependency source with a file and line citation
- A 70-billion-parameter chat model does not fit on one card and must span three
- Four agents in separate worktrees should share one local model
- Files, index and questions must not leave the machine: no telemetry and no account
Not a fit when
- You want a stable release: every PyPI release is a pre-release and interfaces and data formats may change
- You need a cloud model with no local hardware: lilbee is valuable for running on your cards
- The hardware is old with no AVX2: only the `lilbee-compat` build fits and it has no engine wheel for pip
- You need one agent in one folder: plain grep search is enough
Example request
Start Claude Code on a local model with lilbee launch claude and find in the index of my docs folder where authorization is describedMCP
- Transport
- stdio, http
- Authentication
- not required
Limitations
- The project is in active beta: every PyPI release is a pre-release, pip install needs the `--pre` flag, and commands and on-disk formats may change between versions.
- The first `lilbee self-check` downloads about 90 MB and runs a test inference and an embedding.
- Old CPUs with no AVX2 have a separate `lilbee-compat` build, and its pip variant has no engine, so model calls fail without your own `llama-server`.
- macOS and Windows builds are unsigned. ROCm works only on Linux and MI50 cards need the Vulkan build.
How to disable. The package comes off through the manager you used: `brew uninstall lilbee`, `npm uninstall -g lilbee`, `uv tool uninstall lilbee` or deleting the binary. Agent configs where lilbee registered itself as a provider and MCP server are cleaned by hand.
FAQ
How do I install lilbee?
Use a self-contained build: `brew tap tobocop2/lilbee && brew install lilbee`, `npm install -g lilbee` or a binary from the releases. Then `lilbee self-check` and `lilbee`.
How do I run Claude Code on a local model?
Run `lilbee launch claude`. lilbee serves an Anthropic-compatible API and points the session at it through environment variables, leaving Claude Code settings untouched.
What does the agent get from lilbee over MCP?
The `lilbee_search` tool with cited snippets from your library: path, line and text. If the answer is missing, lilbee says so and invents nothing.
Do I need Ollama or a separate server?
No. lilbee has its own model manager and multi-GPU fleet on `llama.cpp`: it downloads models from Hugging Face and runs them on Metal, Vulkan or CUDA.
Does lilbee send data to the cloud?
There is no telemetry or account. lilbee contacts a cloud provider only when you set its API key: once to check the key and then for a cloud model you pick.
Discussion
No questions or reviews yet.
Related
MCP serversllmman runs Claude Code, Codex and OpenCode on a local or hosted model with one command and stores models as OCI images
Local coding agent with a terminal UI: it plans, drafts several candidates, tests them in a sandbox and repairs, so compact open models can handle code
An MCP server that hands bounded tasks to a local model or a cheap cloud API, while Claude stays the architect and saves tokens
A CLI and Python library for OpenAI, Claude, Gemini, local and other models: prompts, chats, tools, embeddings and a SQLite log
A coding agent tuned for small local models, with a minimal system prompt, a planning mode and background sub-agents for research
LlamaForge plugin and MCP server: an agent checks which model fits your GPU, loads it into llama.cpp and hands it routine work
README in short
The README presents lilbee as one program for models and search, warns about beta status, then covers install (bundles, CUDA, ROCm, old CPUs), highlights, sections on launching agents, multi-GPU, the terminal app, the HTTP server, formats and the wiki.