cowork-to-code-bridge — letting a sandboxed agent run work on the developer's own machine

[abhinaykrupa]

DeerFlow's research flow runs wherever you deploy it — a container, a server, a
notebook. That's the right place for orchestration, but it means the agent can't
reach the one machine that has your dotfiles, your SSH keys, your Docker daemon,
and your logged-in CLIs. The usual workarounds are a tunnel or an SSH key handed
to the sandbox, and both widen the trust boundary more than most people want.

I built cowork-to-code-bridge
for that gap. It's deliberately boring: no network listener, no inbound port, no
tunnel. A daemon on your machine watches a shared directory; the sandbox drops a
JSON task file in, the daemon runs it locally and writes a result file back. If
the two sides can see a common folder (bind mount, Syncthing, Dropbox, an NFS
share), they can talk. The machine dials nothing and exposes nothing.

What that buys a DeerFlow-style pipeline:

• A research node can hand a build/test/repro step to the real machine and keep
  planning while it runs — queue_task is non-blocking and returns a task id.
• Execution is allow-listed. The daemon only runs scripts present in its scripts
  directory, so the sandbox can't ask for arbitrary commands.
• Long steps don't wedge the flow: tasks are cancellable (SIGTERM to the whole
  process group, then SIGKILL), output is capped so a runaway build can't blow up
  memory, and every result carries a real exit code.
• Optional per-task plan gate: a task can carry a plan string that a local hook
  approves or rejects before anything executes.

It ships with ~25 small scripts (host health, disk, ports, docker ps/logs, git
status, package updates) plus a run_claude.sh that hands a whole task to Claude
Code on the machine. Python client, MIT, macOS + Linux.

Not trying to replace anything in DeerFlow's execution layer — it's the local
escape hatch for the steps that genuinely need the developer's box. Happy to
answer questions about the design, especially the security model.