What it does
Swarm is a coding agent that runs from the command line against a git repository. You state the change you want in ordinary language, and it modifies your files using either a local or a hosted model, runs the project's own checks, and retries failures a limited number of times. It does not commit: you review the result with git diff and then commit or restore it yourself. The design principle is that checks, not the model's own summary, decide whether work is done. A repair attempt may not delete tests to turn a check green, and a requirement that no trustworthy check covers is reported as not established rather than passed. Each run ends with an outcome such as accepted or partially checked, and prints the path to an evidence bundle that records every command run and can be re-verified offline. It needs Node 22 or newer, git and a repository whose tests run with one command. The page was verified with a Qwen model served through Ollama, which takes about 24 GB of memory, and says the npm release predates what the page describes, so you should build from source. Commands run on your machine under a path policy rather than a sandbox unless you pass an isolation option. The project calls itself experimental and reports no advantage over OpenCode in its latest comparison on public tasks with one local model.
Capabilities
- Plain-language change requests applied to a git working tree
- Runs the project's real checks and repairs failures within a fixed number of attempts
- Repairs cannot delete tests to turn a check green
- Signed, hash-chained evidence record that can be verified offline
- Works with local models through Ollama or with hosted models
When to use it
- You want completion judged by your test suite rather than the model's own report
- You need an auditable record of the commands an agent ran
- You want to try a coding agent driven by a local model
Alternatives to explore
- OpenCode
