Ultraswarm plugin
An orchestrator that hands a coding task to CLI workers in separate Git worktrees, reviews and merges the result.
- 83 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Workers run external CLI agents in parallel that write files and execute commands in worktrees
- Merging edits changes your branch, and tokens and cost grow with the number of workers
- Quality control runs `claude` on your login or API key
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/ultraswarmDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Workers run external CLI agents in parallel that write files and execute commands in worktrees
- Merging edits changes your branch, and tokens and cost grow with the number of workers
- Quality control runs `claude` on your login or API key
This is third-party code. Review the repository files before installing.
What it does
Ultraswarm is a standalone Node runner: it decomposes a task, routes workers, supervises processes, keeps isolated Git worktrees, runs adaptive review and merges changes transactionally with approvals and crash recovery. The Codex, Claude Code, Cursor Agent and Grok hosts stay a thin shell: they prepare a plan and pass it to the runner and write no code themselves. Workers are the CLIs you have installed: `codex`, `gemini`, `grok`, `agy`, `droid`, `opencode`, `pi` and `agent`. At least two with working authentication are required. `claude` is the default brain for review and decomposition. A run report ends with a table of estimated against actual token use by CLI, model and effort. `ultraswarm replan <runId>` emits a plan of failed and blocked tasks, and `ultraswarm add-cli` onboards a new CLI in one command.
Who it is for. Developers who already pay for several coding agents and want to parallelize a large task without sorting branches by hand.
Good fit when
- A large feature can be cut into independent parts and handed to different agents at once
- Edits from parallel agents must not clash, so each needs its own worktree
- A run partly failed and the remaining tasks should quickly become a new plan with `replan`
- You want to see how many tokens each CLI and model actually used
Not a fit when
- The task is small or linear: decomposition and review costs exceed the benefit
- Only one coding CLI is available: at least two authenticated workers are required
- The project is not in a Git repository: isolation relies on worktrees
Example request
/ultraswarm Split the API migration into three independent parts, run them on two workers in separate worktrees and show the report with token use.Limitations
Git, Node 22 or newer and at least two worker CLIs with working authentication are required. By default decomposition and quality control need a `claude` login, or set `ANTHROPIC_API_KEY` and `ULTRASWARM_BRAIN=anthropic-api`. Token use for a CLI without structured reporting is not backfilled: the table shows a dash. External worker cost is bounded through `maxCostUsd` at configured prices. Run state lives in `.ultraswarm/state.sqlite` inside the project.
How to disable. In Claude Code run `/plugin uninstall ultraswarm@ultraswarm`. For Codex and Cursor delete the `~/.agents/skills/ultraswarm` and `~/.cursor/skills/ultraswarm` links. Remove the `.ultraswarm/` directory and created Git worktrees in the project separately.
FAQ
How do I install Ultraswarm in Claude Code?
Run `/plugin marketplace add fubak/ultraswarm`, then `/plugin install ultraswarm@ultraswarm` and invoke `/ultraswarm`. The runner needs a clone of the repository and `npm install`.
How do I install in Codex or Cursor?
In Codex run `bash scripts/install-codex-skill.sh` and invoke `$ultraswarm`. In Cursor use `bash scripts/install-cursor-skill.sh`. The scripts link the skill into `~/.agents/skills/` and `~/.cursor/skills/`.
Which CLIs can be workers?
`codex`, `gemini`, `grok`, `agy`, `droid`, `opencode`, `pi`, `pi-local`, `small-harness` and `agent` are supported. Add your own with `ultraswarm add-cli <name> --binary <bin>`.
What if some tasks failed?
Run `ultraswarm replan <runId>`: it emits a JSON plan of failed and blocked tasks that feeds straight into `run --plan-file -`.
How do I check worker readiness?
`doctor` shows worker state and track record, and `doctor --models` prints the resolved model per CLI and tier.
Discussion
No questions or reviews yet.
Related
PluginsA local Node.js orchestrator that runs Codex, Claude Code, Gemini and OpenCode as one team led by an operator agent
A process kit for agents: a skill router, project memory in a DEV folder and result verification before saying done.
One runtime for Claude Code, Codex, OpenCode and more: a run counts as done only when checks the agent did not write pass
A skill that splits a task across twelve agents such as Codex, Claude Code and Cursor, sharing context and merging results.
Orchestrator plugin for OpenCode and Codex CLI: multiple agents, hooks, built-in MCPs, ultrawork mode and hash-anchored edits
An AI-agent workspace with parallel worktrees, cross-provider orchestration, and bots across chat channels.
Comparison with similar tools
Cards that share the main search query, and how they differ.
README in short
The "What is new" section for version 3.6.0 covers the closing token table, the resolved model recorded per attempt, declared real usage capture and the `replan`, `add-cli` and `doctor --models` commands. Installation for Codex, Claude Code, Grok Build, Cursor Agent and shell, prerequisites and the readiness check follow.