Run coding agents in sandboxes with no GitHub access
TaskWing gives every agent its own sandbox and never gives a sandbox your GitHub login. Your code goes in as a git bundle, only commits come out, and your own machine does the push and opens the pull request. TaskWing refuses to start an agent in a box that holds GitHub credentials. Some of these guards are partial today, and this page says which.
The problem
A coding agent on your laptop runs with your shell, your files and your logins. That is fine when you watch every step. It is not fine when several agents run at once while you do something else. The useful question is not "is the agent careful?" but "what can it reach if it is not?"
What TaskWing does
| Guard | How | Status |
|---|---|---|
| One sandbox per agent | Each piece of the plan gets its own box, forked from a base with your toolchain | Built |
| Code in, commits out | The repository goes in as a git bundle; commits come back as a bundle, over plain exec | Built |
| Only your machine pushes | No box ever gets a GitHub login; the host pushes the branch and opens the draft pull request | Built |
| Refuse a box with GitHub credentials | Agent boxes are checked before an agent starts; a check that cannot run also refuses | Partial: the exam box is not checked yet |
| Secret scan before a box is destroyed | Lists anything that looks like a secret; the agent's own login is named as expected | Partial: a scan that cannot run is skipped silently |
| Hidden tests out of reach | The exam's tests never reach a worker or the lead, and run in a clean box | Built |
| Key proxy for API keys | A placeholder key in the box; the real key is swapped in outside it | Planned |
| Read worker commits safely | Fetch bundles with git hooks and fsmonitor off | Planned |
The full list, with the same statuses, is public in FEATURES.md.
Setting up a clean base
On boat, the default environment passes your GitHub token and secret files into every box. TaskWing refuses that. So you make a separate environment, for example tw-agents, that carries only the agent's login, and build a base box from it:
boat new --environment tw-agents # note the box id
boat ssh <box-id> # install your toolchain; do not log in to gh
boat stop <box-id> # the stop takes the snapshot forks start from
taskwing config boat-base <box-id>
The setup guide has the details, including why a box with no environment at all cannot work (no agent login).
Why open source matters here
The layer that runs your agents decides what they can touch. You should be able to read every line of it, and change it. TaskWing is AGPL-3.0, so copies stay open even when someone runs it as a service.
Questions
Why keep GitHub credentials out of the sandbox?
An agent with a GitHub token can push to any branch it can reach, open pull requests and read other repositories. Without one, the worst it can do is write bad commits, which the exam and your review catch.
How does the agent log in to its model?
Through the sandbox provider. On boat, a separate environment carries only the agent's own login. TaskWing never holds that login. A key proxy for API-key models is planned.
What does local mode isolate?
Nothing. --provider local runs agents in folders on your machine with your own login. It exists to try the flow and for tests, not for untrusted work.
Which sandboxes are supported?
boat today, plus local folders. E2B and AWS are next, and Docker after them. TaskWing never builds its own sandbox.
Does TaskWing scan for leaked secrets?
Yes, every box is scanned before it is destroyed, and the agent's own login is named as expected. The scan is partial: if it cannot run, it is skipped without a warning today.
Try the proof of concept
TaskWing is open source (AGPL-3.0) and early. Today it runs Claude Code agents in boat sandboxes, or in local folders to try the flow.
go install github.com/josephgoksu/TaskWing/cmd/taskwing@main
taskwing run "add a /health endpoint with a test" --provider local --no-pr