swarm-orchestrator CLI
swarm-orchestrator does a code task with a local model, checks each requirement and returns a patch with signed evidence.
- 114 ★ on the repository
- Last commit
Extended access
Why this level
- The agent runs shell commands and edits files from a local model output, without container isolation in `--host` mode.
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/swarm-orchestratorDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent runs shell commands and edits files from a local model output, without container isolation in `--host` mode.
This is third-party code. Review the repository files before installing.
What it does
A local terminal coding agent. `swarm run "Add a --json option to the status command"` reads the task and the repository `AGENTS.md` or `CLAUDE.md`, records requirements with stable IDs and edits files through replacement blocks. Work happens in a clone beside the repository, and every command and edit goes through one chokepoint with a path policy and a ledger. A separate verifier, Swarm Verify, checks the result by running tests in a container with the network closed. The report marks each requirement as satisfied, failed, unverified or blocked, and the patch comes with a signed attestation. The model comes from local Ollama or an OpenAI-compatible server.
Who it is for. Developers who want an agent on a local model without the cloud and a report showing what is done and what stays uncertain.
Good fit when
- Code cannot leave for an external cloud service.
- You need a patch with a list of requirements and evidence for each.
- A task should be replayed and checked with the independent `swarm-verify replay`.
- A local Ollama model exists and should do the edits.
Not a fit when
- There is no Docker, so isolated verification is unavailable and only `--host` remains.
- The local model is weak and does not follow the shell protocol.
- You need a skill for Claude Code, not a standalone agent.
Example request
Run `swarm run` with the task "add a --json option to the status command and keep the text output" and show the requirements report.Limitations
It needs Node 22 or newer, git and a Docker daemon for isolated verification. `--host` runs checks on your machine, and the report marks that as "host, not isolated". The signature shows the report was not altered, but does not prove the code is right or the tests complete. The agent applies the patch to your working tree at the end only if the tree did not change meanwhile. The default branch is `v13-main`, with commits from November 2026.
How to disable. Remove the package with `npm uninstall -g swarm-orchestrator`, plus `swarm-verify` if installed, and the `~/.swarm` folder with model profiles. Work clones stay beside repositories, delete them by hand.
FAQ
Which model does it use?
Without a named model Swarm asks Ollama what is installed, drops cloud entries and picks the strongest one that follows its protocol. Name your own with `--model` or `--endpoint`.
How do I check a model fits?
`swarm doctor --models` shows a verdict for every installed model.
Can I skip applying the patch?
Yes, `--no-apply` keeps the result as a patch.
What does --review do?
The same model in a fresh context looks for one case the change gets wrong and a command that shows it. It is off by default and costs one more model turn.
How do I recheck a result?
Install `swarm-verify` and run `swarm-verify replay <attestation directory> --workspace <repo>`.
Discussion
No questions or reviews yet.
Related
CLI toolsA terminal AI coding agent built on opencode: any model with your own keys, subagent swarms, goals and output compression.
Agent Swarm operator skill
Agent Swarm
Agent Swarm is an operator skill that explains retaining workers’ personal volumes so their identities survive a restart.
An orchestrator that hands a coding task to CLI workers in separate Git worktrees, reviews and merges the result.
A skill that splits a task across twelve agents such as Codex, Claude Code and Cursor, sharing context and merging results.
One runtime for Claude Code, Codex, OpenCode and more: a run counts as done only when checks the agent did not write pass
A team of AI agents from one request: it plans the task, writes code in parallel branches, tests, reviews and opens a pull request
README in short
The README describes the six steps of a run, installation, model selection, work in a clone, the verifier role and the limits of the evidence. It ends with links to ADRs and a comparison with Claude Code.