Ollama Mcp Go

MCP server for Ollama. Lets Claude Code, Cursor, or any MCP client talk to your local Ollama models.

One binary. Zero dependencies. Runs over stdio.

Why this one

There are several Ollama MCP servers out there. This one exists because:

• Go, single binary — no npm, no pip, no runtime. Download and run.
• Agent-oriented — exposes generate (with system prompts), chat (multi-turn), and embed. No admin tools (pull/push/delete) cluttering the tool list.
• Designed for agent-mesh — works as an upstream MCP server behind policy, tracing, and approval workflows. Also works standalone with any MCP client.
• Watchable — MCP delivers a tool result in one piece, so a long reply is invisible until it is done. This server streams from Ollama and mirrors every token to a trace file as it arrives, reasoning included. tail -f and you see the model write.

Tools

Tool	Description
list_models	List all available Ollama models
generate	One-shot generation with optional system prompt
chat	Multi-turn conversation (system/user/assistant messages)
embed	Generate embeddings (for semantic search, RAG, etc.)

generate and chat stream internally and mirror their output to a trace file — see Watching a reply as it is written. The MCP client still receives the reply in one piece, along with the prompt_eval_count / eval_count usage counters.

Conversation state lives entirely on the client: /api/chat is stateless, so every call sends the whole history. Nothing accumulates server-side, and nothing needs clearing here.

Install

From source

git clone https://github.com/KTCrisis/ollama-mcp-go.git
cd ollama-mcp-go
go build -o ollama-mcp-go .

Requires Go 1.22+ and a running Ollama instance.

Pre-built binaries

Coming soon.

Setup with Claude Code

Add to your Claude Code MCP config (~/.claude/settings.json or project .mcp.json):

{
  "mcpServers": {
    "ollama": {
      "command": "/path/to/ollama-mcp-go"
    }
  }
}

If Ollama runs on a non-default host:

{
  "mcpServers": {
    "ollama": {
      "command": "/path/to/ollama-mcp-go",
      "env": {
        "OLLAMA_HOST": "http://192.168.1.10:11434"
      }
    }
  }
}

Setup with agent-mesh

mcp_servers:
  - name: ollama
    transport: stdio
    command: /path/to/ollama-mcp-go

policies:
  - name: agents
    agent: "*"
    rules:
      - tools: ["ollama.*"]
        action: allow

Usage examples

Once connected, your MCP client can call:

Generate with system prompt:

{
  "name": "generate",
  "arguments": {
    "model": "qwen3:14b",
    "prompt": "Explain service meshes in 2 sentences.",
    "system": "You are a concise technical writer."
  }
}

Multi-turn chat:

{
  "name": "chat",
  "arguments": {
    "model": "llama3:8b",
    "messages": [
      {"role": "system", "content": "You answer in French."},
      {"role": "user", "content": "What is the capital of Germany?"}
    ]
  }
}

Embeddings:

{
  "name": "embed",
  "arguments": {
    "model": "nomic-embed-text",
    "text": "AI agent governance"
  }
}

Configuration

Env variable	Default	Description
OLLAMA_HOST	http://localhost:11434	Ollama API URL
OLLAMA_MCP_TRACE	/tmp/ollama-mcp-trace.log	Where replies are mirrored as they stream. Set to off to disable.

Watching a reply as it is written

generate and chat stream from Ollama, but MCP has no way to deliver a partial tool result: the client receives the reply in one piece, at the end. So the server mirrors every token to a trace file the moment it arrives.

tail -f /tmp/ollama-mcp-trace.log

Each call writes a header, the prompt, the reply as it lands, and a footer with the token counts:

=== 21:15:33  chat  gpt-oss:20b ===
> Dis bonjour en trois mots exactement.
---
--- pense ---
Three words greeting: "Bonjour à tous" - Bonjour(1) à(2) tous(3)...
--- repond ---
Bonjour à tous.
[done: 75 prompt / 201 eval]

Models that emit a separate reasoning channel (gpt-oss and friends) have it mirrored under --- pense ---. Reasoning never reaches the MCP client — it goes to the trace and nowhere else, so the answer stays clean while the deliberation stays visible.

Failing to open the trace file is not fatal — tracing turns itself off and generation carries on.

Protocol

Implements Model Context Protocol (MCP) over stdio transport using JSON-RPC 2.0. Compatible with protocol version 2024-11-05.

License

MIT