MCP

lynox implements the Model Context Protocol (MCP) — an open standard for connecting AI tools. Supports stdio and SSE (HTTP Server-Sent Events) transports. This lets you use lynox’s memory, knowledge, and capabilities directly from your IDE or AI assistant.

What This Enables

• Run lynox tasks from Claude Desktop or Cursor without switching windows
• Access lynox’s persistent memory from any MCP-compatible client
• Use lynox as a backend for autonomous task execution

Available Tools

When connected as an MCP server, lynox exposes these tools:

Tool	Description
lynox_run	Run a task and wait for the result
lynox_run_start	Start a task asynchronously (non-blocking)
lynox_poll	Poll for results from an async task
lynox_reply	Answer a question from a running task
lynox_abort	Stop a running task
lynox_memory	Read agent memory by namespace
lynox_reset	Clear a session
lynox_batch	Submit batch tasks for reduced-cost processing
lynox_status	Check batch processing status
lynox_read_file	Read a file from the workspace

Setup with Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on Mac):

Local install:

{
  "mcpServers": {
    "lynox": {
      "command": "npx",
      "args": ["@lynox-ai/core", "--mcp-server"]
    }
  }
}

Docker:

{
  "mcpServers": {
    "lynox": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "ANTHROPIC_API_KEY=sk-ant-...",
        "-v", "~/.lynox:/home/lynox/.lynox",
        "ghcr.io/lynox-ai/lynox:latest", "--mcp-server"
      ]
    }
  }
}

Setup with Cursor

Add to your Cursor MCP settings:

{
  "mcpServers": {
    "lynox": {
      "command": "npx",
      "args": ["@lynox-ai/core", "--mcp-server"]
    }
  }
}

Tool Whitelist

By default, all tools are exposed. To restrict which tools are available:

{
  "mcp_exposed_tools": ["lynox_run", "lynox_memory"]
}

Only listed tools will be registered with the MCP client.

Async Tasks

For long-running tasks, use the async flow:

1. lynox_run_start — Starts the task, returns a run_id immediately
2. lynox_poll — Check for accumulated results (text chunks, tool calls, events)
3. lynox_reply — If the task asks a question, send the answer

This prevents timeouts for tasks that take more than a few seconds.

Tips

• lynox runs tasks with the same capabilities as the Web UI — memory, tools, integrations
• Each MCP session gets its own isolated context
• Use lynox_memory to give your IDE access to business knowledge that lynox has accumulated