How do I install the Swarm skill?

Run npx skillmds@latest add guaardvark/swarm in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Swarm safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Swarm?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Swarm free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

guaardvark (@guaardvark) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Swarm by guaardvark — Agent Skill

Name: Swarm
Author: guaardvark

Swarm

Launch and watch Guaardvark's Swarm Orchestrator: parallel coding agents, each in its own git worktree, working a markdown plan and merging back deterministically. Use when the user wants several agents on one codebase at once, says "swarm this", or asks about a running swarm, worktrees, or merge status.

guaardvark Updated Sep 21, 2026

File contents

Swarm with Guaardvark

Read setup first; needs the swarm plugin (POST /api/plugins/swarm/start). B=${GUAARDVARK_URL:-http://localhost:5000}. Status is also the MCP tool swarm_status.

1. Write the plan file

A markdown file in the target repo. Each ## heading is one task. Structured fields are optional; the parser infers file scope from paths in the prose when they are absent:

# Plan: split the settings page

## Extract the GPU panel
- files: frontend/src/pages/SettingsPage.jsx, frontend/src/components/settings/GpuPanel.jsx
- backend: claude
Move the GPU section into its own component with the same props.

## Add a test for the panel
- files: frontend/src/__tests__/GpuPanel.test.jsx
- depends_on: Extract the GPU panel

backend / assign to picks the agent backend (Claude Code, or a local Cline/OpenClaw agent on Ollama in Flight Mode). Tasks that touch the same files should depend on each other; the parser warns about conflicts.

2. Launch

curl -s -X POST $B/api/swarm/launch -H 'Content-Type: application/json' -d '{
  "plan_path": "/abs/path/repo/docs/plan.md", "repo_path": "/abs/path/repo",
  "max_agents": 4, "auto_merge": false, "flight_mode": false, "dry_run": false,
  "acknowledge_dirty_tree": false
}'

A dirty working tree is refused unless acknowledge_dirty_tree: true; tell the user what is uncommitted first.
self_code: true targets Guaardvark's own checkout and is only allowed on the configured root.
dry_run: true parses the plan and reports tasks and conflicts without starting agents.

3. Watch, merge, clean up

call
`GET $B/api/swarm/status` and `GET $B/api/swarm/status/<swarm_id>`	tasks, states, worktrees
`GET $B/api/swarm/<swarm_id>/logs/<task_id>?lines=100`	one agent's log
`GET $B/api/swarm/<swarm_id>/diff/<task_id>`	its diff before merge
`POST $B/api/swarm/<swarm_id>/bus/broadcast`	message every agent
`POST $B/api/swarm/merge {"swarm_id": "...", "repo_path": "..."}`	merge finished branches in dependency order
`POST $B/api/swarm/cancel`	stop
`POST $B/api/swarm/cleanup {"swarm_id": "...", "delete_branches": false}`	remove worktrees
`GET $B/api/swarm/templates`	saved plan templates

Rules

Review diffs before merge unless the user asked for auto_merge.
Keep max_agents at or below the machine's capacity; each Claude agent is a separate session and each local agent competes for the GPU with generation jobs.
Report per task: state, files touched, whether it merged. Never say "done" on a swarm that still has running or failed tasks.

---name: swarmdescription: Launch and watch Guaardvark's Swarm Orchestrator: parallel coding agents, each in its own git worktree, working a markdown plan and merging back deterministically. Use when the user wants several agents on one codebase at once, says "swarm this", or asks about a running swarm, worktrees, or merge status.---# Swarm with GuaardvarkRead `setup` first; needs the `swarm` plugin (`POST /api/plugins/swarm/start`).`B=${GUAARDVARK_URL:-http://localhost:5000}`. Status is also the MCP tool `swarm_status`.## 1. Write the plan fileA markdown file in the target repo. Each `##` heading is one task. Structured fields are optional;the parser infers file scope from paths in the prose when they are absent:```markdown# Plan: split the settings page## Extract the GPU panel- files: frontend/src/pages/SettingsPage.jsx, frontend/src/components/settings/GpuPanel.jsx- backend: claudeMove the GPU section into its own component with the same props.## Add a test for the panel- files: frontend/src/__tests__/GpuPanel.test.jsx- depends_on: Extract the GPU panel````backend` / `assign to` picks the agent backend (Claude Code, or a local Cline/OpenClaw agent onOllama in Flight Mode). Tasks that touch the same files should depend on each other; the parserwarns about conflicts.## 2. Launch```bashcurl -s -X POST $B/api/swarm/launch -H 'Content-Type: application/json' -d '{ "plan_path": "/abs/path/repo/docs/plan.md", "repo_path": "/abs/path/repo", "max_agents": 4, "auto_merge": false, "flight_mode": false, "dry_run": false, "acknowledge_dirty_tree": false}'```- A dirty working tree is refused unless `acknowledge_dirty_tree: true`; tell the user what is uncommitted first.- `self_code: true` targets Guaardvark's own checkout and is only allowed on the configured root.- `dry_run: true` parses the plan and reports tasks and conflicts without starting agents.## 3. Watch, merge, clean up| call | ||---|---|| `GET $B/api/swarm/status` and `GET $B/api/swarm/status/<swarm_id>` | tasks, states, worktrees || `GET $B/api/swarm/<swarm_id>/logs/<task_id>?lines=100` | one agent's log || `GET $B/api/swarm/<swarm_id>/diff/<task_id>` | its diff before merge || `POST $B/api/swarm/<swarm_id>/bus/broadcast` | message every agent || `POST $B/api/swarm/merge {"swarm_id": "...", "repo_path": "..."}` | merge finished branches in dependency order || `POST $B/api/swarm/cancel` | stop || `POST $B/api/swarm/cleanup {"swarm_id": "...", "delete_branches": false}` | remove worktrees || `GET $B/api/swarm/templates` | saved plan templates |## Rules- Review diffs before `merge` unless the user asked for `auto_merge`.- Keep `max_agents` at or below the machine's capacity; each Claude agent is a separate session and each local agent competes for the GPU with generation jobs.- Report per task: state, files touched, whether it merged. Never say "done" on a swarm that still has running or failed tasks.

Frequently asked questions