cowork-to-code-bridge — let agents running in cloud/sandbox escalate tasks to Claude Code on local machine

[abhinaykrupa]

Hey Anthropic SDK community!

Built something that might be useful for anyone building agents that need to bridge between a cloud/sandbox environment and Claude Code running locally.

The gap it fills

If you're using the Anthropic SDK to build agents, you'll hit this eventually: your agent is in a cloud environment (Cowork, a CI runner, a remote server), but the work needs to happen on a local machine — with access to the user's repos, shell, environment, installed tools.

cowork-to-code-bridge solves this with an async file-based bridge:

from anthropic import Anthropic
from cowork_to_code_bridge.client import queue_task, poll_task_result

client = Anthropic()

# Your agent is running remotely. User needs local work done.
# Queue the task — non-blocking
task_id = queue_task(
    "Build a Flask app with tests, run them, fix failures, push to main",
    idempotency_key="build-flask-app-v1"
)

# Tell the user what's happening
print(f"Delegating to Claude Code on your machine — task {task_id}")

# Later: collect the result
result = poll_task_result(task_id)
if result["status"] == "completed":
    # Use Claude SDK to process the result
    summary = client.messages.create(
        model="claude-opus-4-8",
        max_tokens=1024,
        messages=[{"role": "user", "content": f"Summarize this execution result: {result['output']}"  }]
    )

Architecture

Your Anthropic SDK Agent (cloud/remote)
        ↓ queue_task()
File-based bridge queue (BRIDGE_ROOT/)
        ↓
Local Claude Code daemon (Mac/Linux)
        ↓ runs Claude Code agent locally
Results back via cowork_results/
        ↓ poll_task_result()
Your agent (cloud/remote)

MCP server interface

Also exposed as MCP tools for tool-use workflows:

• escalate_to_claude — queue work, get operation_id
• get_operation_status — idempotent polling with metrics
• cancel_operation — graceful cancellation

Production safety

• Per-operation metrics: tool_calls, api_spend_estimate, memory_mb, cpu_percent
• Loop detection: alerts when same tool called 3+ times (circuit-breaker)
• Budget cap: max_budget_usd per operation
• Graceful cancellation: SIGTERM + timeout
• HMAC token auth, no open ports

Install

curl -fsSL https://raw.githubusercontent.com/abhinaykrupa/cowork-to-code-bridge/main/install.sh | bash

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

Would be curious if anyone here is building hybrid remote/local agent workflows and what patterns you're using for the local execution layer.