DevBox

DevBox is an open-source autonomous engineering harness from RudderStack. You hand it a ticket from Linear or the CLI, and it plans the change, writes the code, runs tests, reviews its own diff and opens a draft pull request that you review and merge. It runs on your infrastructure with your own model keys and is released under the MIT License; the latest listed release is v0.1.0.

What It Is

DevBox is a shared, self-hosted agent harness meant for a whole engineering organization rather than a single developer at a keyboard. It picks up Linear tickets, responds on GitHub pull requests, works across the repositories you list, and keeps learnings from them. The README positions it as working alongside tools like Claude Code and Codex rather than replacing them.

How the pipeline works

Each task moves through orientation, planning, plan review, implementation and self-review, and no code is written until the plan passes review. A control plane stores tasks and starts one sandbox per task through OpenSandbox, without making LLM calls itself. The harness inside the sandbox is built on DeepAgents and carries the task through to a draft pull request. Review comments, @mentions and failed CI runs (with the GitHub App) route back to the running task, so fixes land on the same branch. Learnings are distilled into the repo and committed on the pull request so they are reviewed like code.

Security and model setup

Sandboxes hold only placeholder tokens. An egress proxy called iron-proxy swaps in real tokens and allows GitHub writes only to the repos you configure. Pull request comments are read only from orgs and users on a trust list that is empty by default. Models come from OpenAI or Anthropic, with model and reasoning effort set per stage using your own keys.

Setup path and scope

The quickstart uses Docker Compose or a local kind cluster, Bun, and a GitHub token plus an OpenAI or Anthropic key. Tasks are created and followed with the devbox CLI, and a Helm chart supports Kubernetes. Documented v1 limits: GitHub only, new tasks come from the CLI and Linear, there is no web UI or Jira support, and stopping a task from Linear is not yet supported. RudderStack states that it uses DevBox internally and that it merges about 35% of its internal pull requests with roughly an 80% merge rate.