Description
Claude Code and Codex make one developer faster, but who picks up the team's backlog? DevBox is RudderStack's open source autonomous engineering harness: give it a Linear ticket or a CLI command and, inside your own infrastructure, it plans the change, writes the code, runs the tests, reviews its own diff and opens a draft pull request for you to merge.
Comment on the PR and it pushes the fix to the same branch. Each task gets its own sandbox holding only placeholder tokens, while an egress proxy injects the real ones and limits writes to your repos. Deploy with Docker Compose or Kubernetes, use your own model keys, MIT licensed and written in TypeScript.
Staged pipeline:orientation, planning, plan review, implementation and self-review, with no code written until the plan passes.
Closed feedback loop:review comments, @mentions and failed CI runs go back to the running task.
Versioned memory:learnings are distilled into the repo as a commit on the PR and reviewed like code.
Credential isolation:sandboxes hold dummy tokens and an egress proxy injects the real ones.
Per-stage models:OpenAI or Anthropic, with model and reasoning effort set for each stage.
Comment on the PR and it pushes the fix to the same branch. Each task gets its own sandbox holding only placeholder tokens, while an egress proxy injects the real ones and limits writes to your repos. Deploy with Docker Compose or Kubernetes, use your own model keys, MIT licensed and written in TypeScript.
Features
Staged pipeline:orientation, planning, plan review, implementation and self-review, with no code written until the plan passes.
Closed feedback loop:review comments, @mentions and failed CI runs go back to the running task.
Versioned memory:learnings are distilled into the repo as a commit on the PR and reviewed like code.
Credential isolation:sandboxes hold dummy tokens and an egress proxy injects the real ones.
Per-stage models:OpenAI or Anthropic, with model and reasoning effort set for each stage.
