ollama-mcp

MCP server for delegating coding tasks to local Ollama models.

Setup

cd ollama-mcp
npm install
npm run build

Tools

ollama_list_models

List available Ollama models.

ollama_execute_task

Delegate a coding task to an Ollama model. Returns output with change preview.

• If changes are detected, requires_approval: true is returned

• Review changes, then use ollama_write_files to apply

ollama_write_files

Write files from Ollama output. Use after reviewing changes.

ollama_preview_changes

Analyze text for file write markers and show change preview.

Tool Calling

Models that support tool calling (e.g., qwen3-coder-next, deepseek-r1:8b) use structured JSON tool calls for file writes. This is more reliable than text markers.

The write_file tool is automatically provided to the model:

{
  "type": "function",
  "function": {
    "name": "write_file",
    "description": "Write content to a file at the given path",
    "parameters": {
      "type": "object",
      "properties": {
        "path": { "type": "string" },
        "content": { "type": "string" }
      },
      "required": ["path", "content"]
    }
  }
}

If the model doesn't support tool calling, the server falls back to {file_write} text markers automatically.

Usage

OpenCode

Add to opencode.json:

{
  "mcp": {
    "ollama": {
      "type": "stdio",
      "command": "node",
      "args": ["./dist/index.js"]
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "ollama": {
      "command": "node",
      "args": ["/absolute/path/to/ollama-mcp/dist/index.js"]
    }
  }
}

Workflow

1. ollama_execute_task(task, model)  →  output + changes[]
   - Tool calling models: structured tool_calls
   - Non-tool models: text markers (fallback)
2. Review changes
3. ollama_write_files(output or tool_calls)  →  files written

Environment Variables