Autonomous engineering, shipped by pull request.
Describe a goal. Diffly plans it, runs Claude and Codex agents in isolated sandboxes, gets CI green, merges and deploys, then remembers what it learned.
From goal to deployed
- 1
Plan
Give it a goal. Diffly reads your repo and its memory, then breaks the work into a graph of small, verifiable tasks.
- 2
Build
Claude and Codex agents work in parallel, each in its own gVisor sandbox on its own branch, and never see more than they need.
- 3
Ship
CI, an independent review by a different agent, a preview deploy and staged promotion by digest. Failures roll back.
- 4
Learn
Every outcome feeds a project brain, so the next task starts with your decisions, skills and past mistakes.
Sign in with GitHub, nothing else
Your repositories, your permissions. Diffly only ever touches repos you install it on.
Bring your subscriptions
Agents run on your Claude and Codex plans. No API keys to manage.
Deterministic where it counts
State machines, leases and bounded retries make the control plane predictable. AI plans, builds and reviews.
Reproducible delivery
Images are built once, scanned, tested in a preview, and promoted by digest through staging to production.
Point it at a repo.
Sign in, pick a repository, write a goal.
Get started