About Warren
Warren is open source infrastructure for running coding agents as managed workloads. Instead of an agent living in a developer's terminal session, Warren runs each agent harness as an isolated, observable job on infrastructure the operator controls, and takes care of the surrounding lifecycle.
For each run it creates a fresh worktree or clone on its own branch, keeps the agent inside a sandbox boundary chosen by the operator, and streams live events. It enforces spend and concurrency limits, supports steering and cancellation, and uses watchdogs to recover from lost processes or pods. It also handles Git delivery: agents commit their changes, and Warren manages credentials, branch construction, push and optional pull request creation. Run state, events, cost, token use and outcomes are stored behind one HTTP API, a command line client and a web interface. The guaranteed result of a run is a pushed workspace branch, and project settings can add pull requests, tracker updates and previews.
It is aimed at individual developers and small, trusted engineering teams that want agent runs to continue unattended, repeat on a schedule, survive failures or be visible to others. A small team can share one deployment. The runtime can target Kubernetes.
A public instance at app.warren.run shows a live read only view of Warren building itself. The project is released under the MIT license, and no paid plans are published.