Chat with an AI coding agent that reads, edits, and runs your project's code in a live sandbox. Review the diff, then apply it as a pull request.
Last updated
Agent mode gives each project an AI coding agent with its own live sandbox. Ask it for a change in plain language. It reads your code, edits files, and runs commands like installs, builds, and tests. You review the real diff, and nothing touches your repo until you apply it.
Open your project's Home tab and click Open workspaceon the Agent mode card. If you already have a running session you're taken straight back to it. Otherwise click Start coding. Viewers can open the page, but starting a session, chatting, and applying need the Developer role or higher.
Each session clones your repo into a private sandbox on its own agent/<id>branch. The sandbox comes with Git, Node 22 (npm, pnpm, yarn), Bun, Python 3, Ruby, Go, PHP with Composer, and common build tools. Resources follow the project owner's hosting plan:
| Plan | Memory | CPU |
|---|---|---|
| Free | 768 MB | 0.5 |
| Pro | 1 GB | 1 |
| Business | 2 GB | 2 |
The agent can run shell commands, read files, write files, and list directories inside the sandbox. It follows your project's install, build, and start settings. It tries to verify its own work by running your build, tests, or linter. A single message can use up to 30 tool steps.
Pick a model from the picker in the workspace: Claude, GPT, Qwen, and many more. The default is Qwen3 Coder 30B, which is also the only model available on the free trial. Larger models tend to handle long, multi-step tasks better.
Branch previews follow your plan's branch deployment limit. On plans with none, Apply still pushes the branch, but no preview deployment is created. If an automatic merge is blocked by branch protection or a conflict, the branch stays pushed and you merge it from the PR.