metaswarm plugin
metaswarm runs a task through 23 agents and 9 phases: research, plan, review, TDD execution, PR and a post-merge retrospective.
- 430 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The orchestrator writes code, commits, opens and shepherds pull requests by itself.
- Implementation can be delegated to other CLI agents, each getting access to the repository.
Install
- Works in
Manual install
claude plugin marketplace add dsifry/metaswarm-marketplace
claude plugin install metaswarmRun /setup after install. For Codex: codex plugin marketplace add dsifry/metaswarm-marketplace and $setup. For Gemini CLI: gemini extensions install https://github.com/dsifry/metaswarm.git.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The orchestrator writes code, commits, opens and shepherds pull requests by itself.
- Implementation can be delegated to other CLI agents, each getting access to the repository.
This is third-party code. Review the repository files before installing.
What it does
A self-improving orchestration framework for Claude Code, Gemini CLI and Codex CLI. A user prompt or GitHub issue passes nine phases: research, a plan checked by three reviewers, a design gate with six specialists (PM, architect, designer, security, UX, CTO), decomposition into work units, execution, a final review, PR creation, PR shepherding and closure with knowledge extraction. Each work unit runs an implement, validate, adversarial review, commit loop. The orchestrator validates results itself and does not trust subagent self-reports, and the reviewer is always a different model from the author. Tasks live in BEADS, facts and anti-patterns in a JSONL knowledge base from which agents pull relevant entries before work. Implementation can be delegated to Codex CLI and Gemini CLI to save cost.
Who it is for. Developers and small teams handing agents large tasks who want mandatory gates, TDD and independent review.
Good fit when
- A large feature must be split into work units and each one carried through with tests.
- A plan and design should first pass parallel review by several roles.
- A PR must be shepherded to merge without hand-processing comments and CI.
- Past reviewer feedback should reach the context of future tasks.
Not a fit when
- A one-line fix or quick experiment: mandatory gates will outweigh the task.
- A minimal set is wanted without many agents and mandatory stages.
- The project does not use Git and pull requests.
Example request
/start-task Add a webhook system with retry logic, signature verification and a delivery log UI.Limitations
Per-role models need Claude Code 2.1.251 or newer. The framework imposes a heavy process: 100% test coverage and mandatory TDD. Some features need external tools: BEADS for tasks, Codex CLI and Gemini CLI for delegation, Playwright for visual review. The recommended install is the plugin, while the cross-platform installer `npx metaswarm init` runs third-party code. The autonomous PR Shepherd monitors CI and comments itself and may change the branch.
How to disable. In Claude Code remove the plugin through /plugin. Delete the `.metaswarm` files and copied `.claude/agents/metaswarm-*` agents in the project by hand.
FAQ
How do I install metaswarm in Claude Code?
Run `claude plugin marketplace add dsifry/metaswarm-marketplace`, then `claude plugin install metaswarm`. After that run `/setup`.
Does it work in Codex and Gemini CLI?
Yes. For Codex add the `dsifry/metaswarm-marketplace` marketplace and run `$setup`; for Gemini CLI install the extension with `gemini extensions install`.
What is the Design Review Gate?
A parallel review of the plan by six agents: PM, architect, designer, security, UX and CTO. The gate reruns only after blocking findings, with a cap of 10 rounds.
How do I assign models to roles?
Edit `.metaswarm/models.yaml` and run `/update`. By default reviewers and the planner run on opus, the researcher and coder on sonnet.
How do I see spend by role?
The `metaswarm-usage` command reads Claude Code transcripts and reports subagent cost and review gate rounds.
Discussion
No questions or reviews yet.
Related
PluginsA workflow template for Claude Code and five more agents: a backlog task, a diff review and a commit.
OpenCode plugin that turns a session into a team of agents led by an architect and runs the work through plan phases
A 33-skill plugin with gates: question the request, verify facts against sources, test first, review by a second model
A 13-skill delivery workflow for coding agents: from idea to plan, review, security audit, tests and a shipped pull request
A Grid Dynamics plugin that loads a team's shared engineering standards, architecture and workflows into a coding agent session
A plugin for Codex and Claude Code that runs a task through a chain of roles from planning to an independent audit
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| metaswarm (this card) | no | 430 | |
| Claude Squad | no | 8,790 | manages parallel agent sessions; metaswarm sets phases, gates and a learning knowledge base |
| OpenCode Swarm | no | 526 | a swarm for OpenCode; metaswarm runs in Claude Code, Gemini CLI and Codex with cross-model review |
README in short
The project is extracted from a production orchestration system. It runs recursive orchestration: a swarm coordinator spawns issue orchestrators, which spawn sub-orchestrators for complex epics. After a PR merge the system extracts patterns and anti-patterns from review feedback and build failures on its own.