Building a Model Context Protocol Server for Ollama

Local models are useful. Connecting them to tools is the hard part.

This essay walks through building an MCP server for Ollama: stdio and HTTP transports, how tools should be shaped for agents, and deploying the server so Claude (or any MCP client) can call a local LLM without cloud lock-in.

Takeaways

• Treat tool schemas as product surface — descriptions are the UX for agents
• Stateless HTTP is easier to operate than long-lived stdio in production
• Railway (or similar) works for an always-on bridge between clients and Ollama

If you are exploring MCP for product work, start with a thin read-only server, then add write tools carefully.

Originally published on Medium; this on-site copy is the canonical version.