rudderlabs/devbox

DevBox is worth a look if you want an autonomous coding agent that runs on your machines with your model keys, rather than per-developer sessions in a vendor's environment. The 15-minute quickstart is the real gate: Docker, Bun and an LLM key are prerequisites, and most of that time is the first image build.

The problem

Today an engineer picks up a ticket, switches into a coding assistant, works through the change in one repo, then opens a pull request and waits for CI and review. DevBox changes the entry point: a Linear issue or CLI command becomes the task, and the agent orients itself, writes a plan, implements it, runs tests, reviews its own diff and opens a draft PR. Review comments, @mentions and failed CI runs route back to the same running task, which pushes to the same branch. The engineer's job moves from producing the diff to reviewing and merging it.

What's different

The design is a staged pipeline with a review gate rather than a single agent loop: no code is written until the plan passes. That makes the planning step inspectable and gives the agent a place to fail cheaply. A second distinct piece is credential isolation: the sandbox holds placeholder tokens, and an egress proxy swaps in real ones, permitting GitHub writes only to the repos you list and LLM calls only to your endpoint. Your code stays in the sandboxes and on GitHub; whatever the model needs to see goes to the model endpoint you configure, which is the same exposure you accept with any hosted coding tool.

Memory is versioned: learnings are distilled into the repo and land as a commit on the pull request, so they are reviewed like code rather than quietly accumulating in a private store. Per-stage model routing (OpenAI or Anthropic, with model and reasoning effort set per stage) lets you spend a stronger model on planning or review. The trade-off is operational weight: a control plane with Postgres, sandboxes started through OpenSandbox, and an egress proxy are components you now run and keep running. This is a service for a team, not a local CLI.

Who should try it

A good fit is an engineering team already working in Linear and GitHub that wants background agents handling routine tickets and has someone comfortable running Docker Compose or Kubernetes. The GitHub App path adds a bot identity, self-refreshing tokens and @mention follow-ups, but requires a public HTTPS URL for Linear webhooks and the App, so teams without a tunnel or public endpoint stay on the CLI path with a personal token, and PRs then appear under that user's name. A poor fit is a solo developer who wants an interactive pair-programming experience, or anyone unwilling to run a persistent control plane alongside their repos. The project's own claim of roughly 35% of RudderStack's internal PRs and an 80% merge rate is the author's account of their internal rollout, not an independent benchmark.

Requirements and limitations

The quickstart assumes macOS (with notes for Linux) and needs Docker Desktop with 8 GB of memory for Docker, Compose 2.13+, Bun 1.4.4+, make, openssl, a GitHub token or logged-in gh, and an OpenAI or Anthropic key. Config restricts writes to the repo URLs you list. Comment-driven follow-ups only read from orgs and users you explicitly trust, and that list is empty by default, so the closed feedback loop does not work until you fill it in. Merging stays with you and branch protection still applies.

The bigger ongoing cost is infrastructure: control plane, Postgres, sandbox runtime and proxy need to stay healthy, and Linear/GitHub App integration needs a stable public URL. Secrets are kept out of config files, which shifts care toward the keychain and environment where they live. See the architecture notes and the configuration details for the moving parts.

HotRepo's take

HotRepo recorded the repository at 122 total stars, ranked #11 on the NEW-REPOSITORY board for TypeScript on 2026-10-28. That single dated snapshot, for a repository one day old, cannot show whether adoption persists; comparable daily history is insufficient. Before investing setup time, check the trust settings and proxy allowlist in the configuration docs, since those govern what the agent can actually touch.