Workspaces: your workflows can write code

Until today, a Murmurator workflow could read a pull request, reason about it and comment on it — but it couldn't change a line. Workspaces close that gap. A run can now clone a repository into a folder of its own, edit files, run your test suite and push a branch.

What's new

• A folder per run. workspace.create provisions one, optionally cloning a repository through a GitHub connection. It lives in a Docker sandbox and is deleted, with everything in it, the moment the run finishes.
• A coding agent. workspace.agent runs a coding agent over the folder: it reads the files, edits them and runs commands on its own, in a loop. Describe the job the way you'd describe it to a developer — “the rounding test fails, find the cause and fix it” — and it goes and does it. Pick which one on workspace.create: opencode by default, or codex.
• Your tests, not its word. workspace.exec runs anything the repository builds with and hands back the exit code, so the rest of the workflow can branch on a suite that actually passed. workspace.diff returns the patch for a person, or for an llm step, to read.
• Straight into review. workspace.commit and workspace.push put the work on a branch, and the new github.create_pull_request tool opens the pull request.
• Nothing to set up. Workspace tools are there for every account, like the built-in tools, with no connection to add. A GitHub connection is only needed for a private repository or for pushing.

Where it's kept in

A coding agent is a big thing to hand an automation, so it is boxed in on every side. It works inside a Docker sandbox with no route to our servers and none of your credentials: paths that resolve outside its folder are refused, the GitHub token is handed to git for the one command that needs it and is gone before the agent runs, and cloning and pushing obey that connection's allowed repositories.

The agent needs one of your own AI models. opencode takes any provider you can add, including a model you host yourself on Ollama; codex speaks only the OpenAI responses API, so it takes an OpenAI or Ollama model. Murmurator AI models aren't accepted here, because the agent calls the provider directly and we would have no honest way to meter what it spends.

One thing worth keeping in mind: a coding agent reads the repository, and repositories are written by other people, so a README or an issue template can try to talk it into something. Give the connection it pushes with an allowed-repositories list, push to a branch rather than to main, and keep a person on the merge button.

See the Workspaces guide, or the issue-to-pull-request example for the whole thing end to end.