Install The Skills
This repo ships these installable skills:
| Skill | Purpose |
|---|---|
skills/swarm-mcp | General swarm coordination plus role workflows for planner, implementer, reviewer, researcher, and generalist sessions |
skills/swarm-deepdive | Forensic inspection of a swarm — reconstruct timelines, debug stuck agents, audit task/message/KV history straight from swarm.db and the server logs |
Invoke swarm-mcp directly with a role argument when you want a session to adopt a role immediately:
/swarm-mcp planner
/swarm-mcp implementer
/swarm-mcp reviewer
/swarm-mcp researcher
/swarm-mcp generalist
Invoke swarm-deepdive when you want a session to investigate what a swarm did rather than participate in one — postmortems, debugging stuck tasks, auditing message/KV history.
Each skill's main SKILL.md stays short and routes to supporting reference files under skills/<name>/references/ only when needed.
Important boundary:
- the skills teach agents how to use
swarm-mcpand inspect its history well - the skills do not install or mount the MCP server for you
- configure the
swarmMCP server first, then install the skills
Automated Project Setup
When using the packaged CLI, swarm-mcp init can write a project-local MCP config and copy the bundled skills:
swarm-mcp init --dir /path/to/project
This writes:
.mcp.jsonwith aswarmserver entry that runsnpx -y swarm-mcp.claude/skills/swarm-mcpand.claude/skills/swarm-deepdiveunless--no-skillsis passed
Useful flags:
--force: overwrite an existing copied skill--no-skills: write only.mcp.json--dir <path>: install into a project directory other than the current directory
Use the manual sections below for global installs, non-Claude skill directories, or local-clone MCP configs.
Symlink From A Local Clone (Recommended For Consumers)
If you cloned this repo or vendor it under another project, prefer symlinks over copies. Symlinks keep the skills single-sourced — when you git pull swarm-mcp, every consumer picks the update up automatically.
The skill sources live at:
/path/to/swarm-mcp/skills/swarm-mcp
/path/to/swarm-mcp/skills/swarm-deepdive
Install into a consumer project (matches the convention this repo itself uses for its bundled skills, where .claude/skills/<name> is a symlink to ../../.agents/skills/<name>):
# In the consumer project
mkdir -p .agents/skills .claude/skills
ln -s /absolute/path/to/swarm-mcp/skills/swarm-mcp .agents/skills/swarm-mcp
ln -s /absolute/path/to/swarm-mcp/skills/swarm-deepdive .agents/skills/swarm-deepdive
ln -s ../../.agents/skills/swarm-mcp .claude/skills/swarm-mcp
ln -s ../../.agents/skills/swarm-deepdive .claude/skills/swarm-deepdive
Or globally for every project on the machine:
mkdir -p ~/.claude/skills
ln -s /absolute/path/to/swarm-mcp/skills/swarm-mcp ~/.claude/skills/swarm-mcp
ln -s /absolute/path/to/swarm-mcp/skills/swarm-deepdive ~/.claude/skills/swarm-deepdive
Most agent hosts follow symlinks transparently. If yours does not, fall back to the copy-based instructions below.
Skill Source
Copy each skill folder as a unit:
skills/swarm-mcp
skills/swarm-deepdive
The destination folder must keep the skill name and contain SKILL.md plus its references/ directory. Claude Code treats skills as slash commands, so /swarm-mcp planner is the role bootstrap command and /swarm-deepdive loads the forensic inspection workflow.
Claude Code
Install project-locally by copying the folder to:
.claude/skills/swarm-mcp
.claude/skills/swarm-deepdive
Or install globally for all projects:
~/.claude/skills/swarm-mcp
~/.claude/skills/swarm-deepdive
After installing, invoke /swarm-mcp directly or pass a role argument like /swarm-mcp implementer; invoke /swarm-deepdive for investigations.
Codex
Install globally by copying the folder to:
~/.codex/skills/swarm-mcp
~/.codex/skills/swarm-deepdive
On Windows, that is typically:
C:\Users\<you>\.codex\skills\swarm-mcp
C:\Users\<you>\.codex\skills\swarm-deepdive
PowerShell example:
New-Item -ItemType Directory -Force "$HOME/.codex/skills" | Out-Null
Copy-Item -Recurse -Force "C:\path\to\swarm-mcp\skills\swarm-mcp" "$HOME/.codex/skills\swarm-mcp"
Copy-Item -Recurse -Force "C:\path\to\swarm-mcp\skills\swarm-deepdive" "$HOME/.codex/skills\swarm-deepdive"
Then restart Codex so it picks up the new skill.
OpenCode
OpenCode officially discovers skills in these locations:
- project-local:
.opencode/skills/<name> - global:
~/.config/opencode/skills/<name> - project Claude-compatible:
.claude/skills/<name> - global Claude-compatible:
~/.claude/skills/<name> - project agent-compatible:
.agents/skills/<name> - global agent-compatible:
~/.agents/skills/<name>
A simple project-local install is:
.opencode/skills/swarm-mcp
.opencode/skills/swarm-deepdive
PowerShell example from a repo root:
New-Item -ItemType Directory -Force ".opencode/skills" | Out-Null
Copy-Item -Recurse -Force "C:\path\to\swarm-mcp\skills\swarm-mcp" ".opencode/skills\swarm-mcp"
Copy-Item -Recurse -Force "C:\path\to\swarm-mcp\skills\swarm-deepdive" ".opencode/skills\swarm-deepdive"
OpenCode loads skills on demand through its native skill tool. Restarting the session is still a good idea after adding a new one.
Claude-Compatible Shared Folder
If you want one project-local location that OpenCode already treats as Claude-compatible, use:
.claude/skills/swarm-mcp
.claude/skills/swarm-deepdive
Or globally:
~/.claude/skills/swarm-mcp
~/.claude/skills/swarm-deepdive
This is useful when you already organize reusable agent instructions under a Claude-style skill directory and want OpenCode to discover the same skill.
Verify The Install
After installing the skills and mounting the swarm MCP server:
- Start a fresh session
- Confirm the host can see the
swarm-mcpandswarm-deepdiveskills - Ask the agent to join the swarm or invoke
/swarm-mcp planner - Verify it calls
register, then checkslist_instances,poll_messages, andlist_tasks - Invoke
/swarm-deepdivewhen you want an agent to inspectswarm.dbor server logs
If the skills appear but the swarm tools do not, the skill install worked and the MCP setup did not.
Recommended Setup
Use all three layers together:
- MCP config: makes the swarm tools available
AGENTS.md: gives ambient always-on rulesSKILL.md: gives the agent reusable playbooks it can load when relevant
That combination is more reliable than any one layer by itself.
