Local AI MCP

Unified MCP server for managing local model runtimes (Ollama, LM Studio, and more): provider-agnostic discovery, lifecycle, hardware-fit, and delegated inference.

Local AI MCP is an MCP server that turns your local model runtimes into an agent-callable control plane. It is operations-first: its primary job is to discover, inspect, fit, and manage the models running on your own machine. It speaks to runtimes over their local HTTP APIs and exposes one consistent tool surface across them, so an agent does not need to know whether a model lives in Ollama or LM Studio.

The server communicates over stdio only. It is a client to your local runtimes and never opens a network listener of its own.

Why an ops-first local-model server

• Discovery and lifecycle, not just chat. List what is installed, what is loaded, pull and remove models, load and unload them, and check their fit against your hardware before you commit VRAM to them.

• Hardware-aware. system_resources and fit_check read your real RAM and GPU/VRAM so an agent can pick a model that will actually run, and suggest_model ranks candidates by task and by what fits.

• Provider-agnostic. Every tool takes an optional provider argument. Omit it and the tool operates across all detected runtimes, aggregating results per provider.

Related MCP server: Ollama MCP Server

Inference is delegation, not chat

The complete and embed tools exist to delegate (offload) inference to a local model for cost control and privacy: keep tokens and data on your own hardware instead of sending them to a hosted API. They are deliberately framed as delegated/offloaded inference primitives, not as a conversational chat surface.

The provider-adapter model

Each runtime is implemented as an adapter behind a single Provider interface (src/providers/types.ts) with a uniform method set: detect, health, listModels, listLoaded, modelInfo, pull, remove, load, unload, complete, embed, and capabilities. Adding a runtime means adding one adapter; the tool layer is unchanged.

Auto-detection: on each call the server probes the configured local endpoints to determine which runtimes are live. Hardware probing is isolated in src/hardware/ and branches by platform (Windows / Linux); it exposes total/free RAM and, where detectable, GPU name and VRAM.

Tool surface (16 tools)

Discovery

Lifecycle

Ops

Registry

Delegation (offloaded inference)

Every tool except system_resources accepts an optional provider (ollama | lmstudio). Omit it to operate across all detected runtimes.

Install and run

npx @tmhs/local-ai-mcp

Claude Desktop / Cursor config

{
  "mcpServers": {
    "local-ai": {
      "command": "npx",
      "args": ["-y", "@tmhs/local-ai-mcp"],
      "env": {
        "OLLAMA_HOST": "http://localhost:11434",
        "LMSTUDIO_HOST": "http://localhost:1234"
      }
    }
  }
}

Configuration

All configuration is via environment variables with sane defaults:

Development

npm install
npm run build      # tsc -> dist/
npm test           # vitest; runs fully offline (mocked HTTP, stubbed hardware)

The test suite requires no running runtime and no downloaded model: every HTTP call is mocked and hardware probing is stubbed.

License

CC-BY-NC-ND-4.0 -- see LICENSE.

Built by TMHSDigital