bionic-mcp

You run models in LM Studio. bionic-mcp turns them into tools Claude Code (or any MCP client) can call, so you can hand the cheap and bulky work to your own machine instead of a paid model. It's free, local, and private: the prompt and the answer never leave your Mac.

Why not just point the client at LM Studio's API

You can, and two things will eventually bite you. This server fixes both:

• LM Studio silently swaps models. Ask for a model that isn't loaded and its API quietly answers with whatever is loaded, so you think you're talking to a 14B and you're actually talking to a 3B. This server refuses an unknown model id up front and checks response.model on every reply, so you always know what answered.

• Cold loads are slow. On 16GB only one model stays resident, and switching costs 8-30 seconds. model:"auto" uses whatever's already loaded instead of forcing a reload, so a quick call stays quick.

That's the whole pitch: safe access to the models you already run, without the two foot-guns.

Related MCP server: Local LLM MCP Server

Where it fits your flow

You're in Claude Code and hit a subtask that doesn't need the expensive model: classify these, summarize that, second-opinion this, extract fields from a batch. Hand it to respond and it runs on your hardware for free. models shows you what's installed and loaded so you can pick, or let auto decide.

Install

Claude Code (copy-paste)

Paste this into your terminal. It clones, installs, and registers the MCP in one shot:

git clone https://github.com/turner-moore/bionic-mcp.git && cd bionic-mcp && npm install && \
claude mcp add -s user bionic -- node "$(pwd)/index.js"

You also need LM Studio running its local server (that's where the models are):

~/.lmstudio/bin/lms server start

Any other MCP client

Add this to your client's config (e.g. claude_desktop_config.json) with an absolute path:

{
  "mcpServers": {
    "bionic": {
      "command": "node",
      "args": ["/absolute/path/to/bionic-mcp/index.js"]
    }
  }
}

Point it at a different endpoint with BIONIC_BASE_URL (default http://127.0.0.1:1234).

Tools

• respond: run a prompt. Pick a model by id, or let model:"auto" use the loaded one. Takes a system prompt, temperature, and a JSON schema to force structured output.

• models: list installed models with live load state and a capability card (quality, speed, context size), so you can decide what to hand off.

One rule that matters: each prompt has to stand on its own. The local model can't see your Claude session, files, or memory, so inline every input and instruction it needs.

Two servers in here

• index.js (bionic): the straight LM Studio bridge. Start here.

• local.js (local): an optional tiered version. It tries a faster on-device backend first (caix, Apple Core AI on the Neural Engine) and falls back to LM Studio when that's down. Register it separately if you want it:

  claude mcp add -s user local -- node "$(pwd)/local.js"

Requirements

Node 18+ (for built-in fetch) and LM Studio with its server running. The capability card lives in models.json, edit it to match your own installed models.