Run host-side actions from a sandboxed Burr app via a local-first bridge (async, idempotent)

[abhinaykrupa]

TL;DR: cowork-to-code-bridge is a small, local-first, MIT-licensed bridge that lets an agent running in a sandbox (Claude Cowork, or any MCP-aware framework) hand real work to Claude Code on your own machine — build/run apps, run tests, git push, docker, brew/pip/npm, check RAM/disk/processes — and get results back, all over a shared bind-mounted directory.

Why it might interest Burr users: Burr models apps as state machines with actions; some of those actions legitimately need to run on the host (kick off a build, run a test suite, shell out) while the orchestrating process is sandboxed or remote. The bridge gives you a durable, async boundary for exactly that escape hatch — without inbound ports or a tunnel.

The async shape maps cleanly onto a state machine:

• queue_task() returns a task_id immediately — your action can transition to a "waiting" state instead of blocking
• poll_task_result(task_id) is idempotent — safe to call every tick from a polling node
• call_remote() when you do want a blocking action
• idempotency_key so a re-entered/retried action doesn't double-fire side effects

It's a file-based RPC (a BRIDGE_ROOT dir + a local daemon), stdlib-only on the sandbox side, idempotent, and reboot-surviving.

Repo: https://github.com/abhinaykrupa/cowork-to-code-bridge — happy to hear where this abstraction breaks down for Burr-style workflows.