cowork-to-code-bridge — MCP server that exposes Claude Code on local machine to any MCP client

[abhinaykrupa]

Hey mcp-use community!

Built something that extends the MCP pattern into local machine execution — thought it might be relevant here.

What it is

cowork-to-code-bridge exposes Claude Code (running on your local Mac/Linux) as an MCP server that any client (including mcp-use) can call. Instead of calling a cloud tool, your MCP client escalates a task to Claude Code on the user's actual machine — with full repo, shell, and environment context.

MCP interface

Three tools:

• escalate_to_claude — queue a task, returns operation_id immediately (non-blocking)
• get_operation_status — poll status with full metrics (tool_calls, api_spend, memory, CPU)
• cancel_operation — graceful SIGTERM + idempotent on completion

How it works with mcp-use

from mcp_use import MCPClient

# Connect to the bridge MCP server
client = MCPClient.from_config({
    "mcpServers": {
        "local-claude-code": {
            "command": "cowork-to-code-bridge-mcp",
            "args": ["--stdio"]
        }
    }
})

# Escalate a task to Claude Code on the local machine
result = await client.call_tool(
    "local-claude-code",
    "escalate_to_claude",
    {"prompt": "Build a FastAPI health check endpoint and run the tests"}
)
operation_id = result["operation_id"]

# Poll for completion
while True:
    status = await client.call_tool(
        "local-claude-code",
        "get_operation_status",
        {"operation_id": operation_id}
    )
    if status["status"] == "completed":
        print(status["output"])
        break
    await asyncio.sleep(2)

Production safety

• Per-operation metrics in every get_operation_status response
• Loop detection (circuit-breaker when same tool called 3+ times)
• Budget caps (max_budget_usd per operation)
• HMAC token auth, no open ports
• Graceful cancellation

Repo: https://github.com/abhinaykrupa/cowork-to-code-bridge

Would love to hear if anyone's tried bridging local Claude Code this way through mcp-use. The async pattern (fire-and-forget with polling) is designed to avoid blocking mcp-use's event loop.