Bridging Ollama to MCP: Using MCPHost for Local LLM Tool Calls

Ollama does not implement the MCP protocol; it exposes an OpenAI‑compatible REST API with a /api/chat endpoint and a tools parameter that returns JSON tool calls. To use MCP servers with a local Ollama model, the article recommends MCPHost, a Go CLI that bridges MCP servers to Ollama by negotiating capabilities and routing tool calls. MCPHost is installed via go install github.com/mark3labs/mcphost@latest, configured with a JSON file listing MCP servers (e.g., filesystem) and specifying the Ollama backend URL. The command mcphost --config mcp-config.json --model ollama:qwen2.5:14b --ollama-url http://localhost:11434 starts the bridge, enabling the model to call MCP servers. Supported models that handle tool calling include Qwen2.5, Llama 3.3, Mistral Nemo, and Gemma 3; smaller quantizations may produce malformed JSON. Limitations include tool‑calling reliability, context‑length consumption, and latency, so users should simplify schemas, truncate large responses, and expect 2–10 s per call on consumer hardware. For offline use, stick to filesystem, database, and local code servers, while cloud‑service servers require API credentials. MCPHost is the simplest CLI for local MCP integration, but other MCP clients can also be used.