Run, fit, and route work to local LLMs (Ollama, LM Studio, llama.cpp, MLX, vLLM, and more).
Maintain this server? Claim it
This server only runs locally, from a package. Install it with the instructions on this page to try it.
Pick your client. Values in capitals, like YOUR_TOKEN, are yours to fill in.
claude mcp add local-ai-mcp \ -- npx -y @tmhs/[email protected]
Environment
OLLAMA_HOST (optional, not in the config above): Ollama base URLLMSTUDIO_HOST (optional, not in the config above): LM Studio server base URLLLAMACPP_HOST (optional, not in the config above): llama.cpp server (llama-server) base URLDOCKER_MODEL_RUNNER_HOST (optional, not in the config above): Docker Model Runner host TCP endpointMLX_HOST (optional, not in the config above): mlx_lm.server base URL (opt-in; Apple Silicon)VLLM_HOST (optional, not in the config above): vLLM OpenAI server base URL (opt-in)VLLM_API_KEY (optional, not in the config above): Bearer token for VLLM_HOST (when vLLM runs with --api-key)FOUNDRY_LOCAL_HOST (optional, not in the config above): Foundry Local service URL (opt-in; port from `foundry service status`)OPENAI_COMPAT_HOST (optional, not in the config above): Any OpenAI-compatible /v1 server (opt-in)OPENAI_COMPAT_API_KEY (optional, not in the config above): Bearer token for OPENAI_COMPAT_HOSTMOONSHOT_API_KEY (optional, not in the config above): Moonshot AI (Kimi) key. Hosted: only used when named explicitly.LOCAL_AI_ALLOW_HOSTED_FALLBACK (optional, not in the config above): Allow implicit routing to hosted providers (default: off)MOONSHOT_HOST (optional, not in the config above): Moonshot API base URLLOCAL_AI_DISABLE (optional, not in the config above): Comma-separated provider names to skip probingLOCAL_AI_PROVIDERS (optional, not in the config above): JSON array of extra provider instances: {type, host, name?, apiKey?, kind?}LOCAL_AI_DISABLE_NETWORK_SEARCH (optional, not in the config above): Keep search_models offline (catalog only)LOCAL_AI_BENCHMARK_CACHE (optional, not in the config above): Benchmark results file used to rank delegate by measured speed (default: per-user cache dir; off disables)We don't run packages, so we can't list this server's tools.
Runs on your machine from a package. We don't run packages, so it isn't checked.