Imported from glitch-rabin/swarma (SKILL.md). Install upstream with npx skills add glitch-rabin/swarma. Copyright stays with the author (MIT).

swarma -- growth experiment loop for agent teams

When to Use This Skill

Use swarma when the user wants to:

• Run growth experiments (hooks, landing pages, outreach, pricing, activation, retention)
• Build agent teams that learn and improve through A/B testing, not just execute once
• Get a validated playbook of what actually works for their specific audience/product
• Test ideas at scale (50+ experiments/week instead of 2-5)
• Replace "we tried that, it didn't work" with logged, analyzed, searchable experiment data

Trigger phrases: "test what works", "optimize my funnel", "find the best hooks", "run experiments", "A/B test", "what's working", "build a playbook", "growth experiments", "improve conversion"

Do NOT use when: user wants workflow automation (use n8n/Make), conversation memory (use honcho), or one-shot agent pipelines (use CrewAI/AutoGen). swarma is specifically for experiment loops that improve over time.

Quick Reference

Commands at a Glance

Decision: Which Squad Template?

Decision: Generate vs Template?

The GROWS Loop (Core Concept)

Every experiment cycle follows five steps:

  Generate       Run         Observe       Weigh        Stack
 hypothesis --> experiment --> signal --> verdict --> playbook
     ^                                                  |
     └──────────────────────────────────────────────────┘

Key numbers:

• Verdict threshold: 20% improvement to keep, 20% decline to discard
• Default min_sample_size: 5 cycles before verdict
• Scoring: 1-10 scale with forced decimals (7.3, not 7)

Setup Guide

Platform: Claude Code / Claude Desktop

pip install swarma
swarma init

Add to .mcp.json:

{
  "mcpServers": {
    "swarma": {
      "command": "swarma",
      "args": ["serve", "--mcp"],
      "env": { "OPENROUTER_API_KEY": "sk-or-..." }
    }
  }
}

Important: OPENROUTER_API_KEY must be in the MCP env block. The instance .env is not inherited by subprocesses.

Platform: Hermes (via terminal)

Hermes has terminal access -- it can run swarma CLI commands directly. No MCP required.

pip install swarma
swarma init

Then tell Hermes: "run swarma cycle hook-lab --topic 'AI agents are overhyped'"

Hermes reads terminal output and acts on results. For structured access, add MCP:

# hermes config.yaml
mcp_servers:
  swarma:
    transport: stdio
    command: swarma
    args: ["serve", "--mcp"]
    env:
      OPENROUTER_API_KEY: "sk-or-..."

Platform: OpenClaw

pip install swarma
swarma init

Configure as MCP tool or use terminal access depending on your OpenClaw setup.

Platform: CLI (standalone)

pip install swarma
swarma init                                        # creates instance + starter team
swarma cycle starter --topic "why do startups fail?"   # run one cycle
swarma status                                      # check costs, runs, experiments

From source

git clone https://github.com/glitch-rabin/swarma.git
cd swarma && pip install -e .
swarma init

Environment setup

After swarma init, add your API key:

echo "OPENROUTER_API_KEY=sk-or-..." >> ~/.swarma/instances/default/.env

Get a key at openrouter.ai/keys.

Optional (for cross-team knowledge):

# Only needed when running 3+ teams
echo "QMD_ENDPOINT=http://localhost:8181/mcp" >> ~/.swarma/instances/default/.env

Onboarding Flow

When a user wants to set up swarma, follow this sequence. The team generator is the fastest path -- don't make users configure agents manually.

Step 1: Understand the goal

Ask:

• "What do you want to improve?" (conversion, engagement, outreach response rate, SEO rankings, etc.)
• "Who is your audience?" (B2B SaaS users, crypto community, enterprise buyers, etc.)
• "What does success look like?" (more signups, higher CTR, better reply rates, etc.)

Step 2: Install

pip install swarma
swarma init --yes

Step 3: Generate the team

This is the key step. Use the team generator instead of picking templates.

swarma team create growth-lab \
  --from-goal "optimize landing page conversion for our B2B SaaS" \
  --context "developer tools company, 500 free users, 2% conversion to paid" \
  --budget 30

The generator:

1. Designs the team (2-5 agents with specific roles)
2. Picks models that fit each role
3. Writes agent instructions and experiment patterns
4. Creates a first experiment hypothesis ready to run

Review what it generated:

swarma team show growth-lab

Step 4: Run the first cycle

swarma cycle growth-lab

Expected output:

Running cycle: growth-lab
  flow: researcher -> copywriter -> judge
  agents: ['researcher', 'copywriter', 'judge']

                              Cycle: growth-lab
  Agent      Model              Cost       Output Preview
  researcher sonar-pro          $0.000384  **Topic:** 52% of executives...
  copywriter qwen3.5-plus-02-15 $0.000746  [A] We sent 4,382 cold emails...
  judge      mistral-nemo       $0.000416  **Hook Variations:** A: "Did...

  duration: 43.9s | total cost: $0.001546 | agents: 3

Step 5: Run more cycles and review

swarma cycle growth-lab                    # run another cycle
swarma cycle growth-lab --topic "specific angle"  # with a topic
swarma status                              # check progress

After 5 cycles, the experiment engine issues its first verdict. The strategy file evolves automatically.

Day-to-Day Usage

Running experiments

# Single cycle
swarma cycle hook-lab

# With a specific topic
swarma cycle hook-lab --topic "AI agents are commoditizing"

# Continuous (teams with cron schedules run automatically)
swarma run

# Continuous with API server
swarma run --port 8282

Feeding real metrics

LLM self-eval is a starting proxy. For production, feed back real-world signals:

# Log a single metric
swarma metric log hook-lab copywriter 4.2 --metric ctr_pct

# Attach to a specific experiment
swarma metric log hook-lab copywriter 127 --metric impressions --exp 3

# Add a note
swarma metric log hook-lab copywriter 5.1 --metric ctr_pct --note "from linkedin analytics"

# Bulk import from CSV
swarma metric import hook-lab metrics.csv

# View logged metrics
swarma metric show hook-lab

CSV format: agent,value,metric_name,note

copywriter,4.2,ctr_pct,week 1
copywriter,5.1,ctr_pct,week 2
researcher,7.8,relevance_score,

Using squad templates

# Copy a template to your instance
cp -r "$(python -c "import swarma; print(swarma.__path__[0])")/examples/hook-lab" \
  ~/.swarma/instances/default/teams/hook-lab

# Or if you cloned the repo
cp -r examples/hook-lab ~/.swarma/instances/default/teams/hook-lab

# Run it
swarma cycle hook-lab --topic "why most startups fail"

Checking status

swarma status

Shows: all teams, recent runs, costs (today + this month), pending plans, queue stats.

MCP Tools Reference

When connected via MCP, these 16 tools are available:

Common MCP Workflows

"What's been happening?"

1. swarma_status -- overview
2. swarma_get_outputs -- recent agent outputs
3. swarma_list_plans -- pending experiments

"Run an experiment"

1. swarma_run_cycle with team_id and optional topic
2. swarma_get_outputs to review results

"Start a new experiment area"

1. swarma_generate_team with goal and context
2. swarma_get_team to review what was generated
3. swarma_run_cycle to kick it off

"What's working?"

1. swarma_get_outputs for recent results
2. Read the team's strategy.md for validated patterns

Team Configuration Reference

A team is a folder. No code required.

teams/my-squad/
├── team.yaml          # goal, flow, schedule, budget
├── program.md         # team context and constraints
└── agents/
    ├── researcher.yaml
    ├── writer.yaml
    └── strategy.md    # pre-seeded growth knowledge (evolves automatically)

team.yaml

name: my-squad
goal: find what works.
flow: "researcher -> writer"        # sequential
# flow: "researcher -> [writer, analyst]"  # parallel
schedule: "0 8 * * 1-5"            # optional: weekdays at 8am
budget: 30                          # optional: monthly budget in $

agent.yaml

id: writer
name: Writer
instructions: |
  turn research into a post. max 200 words.
  hook in the first line. practitioner voice.
model: qwen/qwen3.5-plus-02-15     # optional: override default routing
metric:
  name: content_quality
  target: 8.0
experiment_config:
  min_sample_size: 5
  auto_propose: true

strategy.md (evolves automatically)

Starts with seed knowledge, grows with every validated experiment:

### Validated Patterns

**Specificity wins**
- Hooks with specific numbers outperform vague claims by 2-3x on saves
- "47% of startups" > "most startups"

### Anti-patterns (Discarded)
- Generic inspirational openings: -23% vs baseline. Discard.

### Patterns to Test
- [ ] First-person confession vs third-person case study
- [ ] Time-anchored ("In 2024...") vs timeless hooks

Flow DSL

# Sequential: a runs, output passes to b
flow: "researcher -> writer"

# Parallel: a runs, then b and c run concurrently
flow: "researcher -> [writer, analyst]"

# Mixed: sequential then parallel then sequential
flow: "researcher -> [writer, analyst] -> judge"

Cross-Team Knowledge (QMD)

By default, each team learns individually via its own strategy.md. To share knowledge across teams, wire in QMD:

# ~/.swarma/instances/default/config.yaml
knowledge:
  engine: qmd
  qmd_endpoint: http://localhost:8181/mcp

With QMD: team A discovers loss framing beats gain framing, team B sees that pattern in its next cycle. Anti-patterns are shared too.

You don't need QMD until running 3+ teams. Most users start without it.

Troubleshooting

Verification

After setup, verify everything works:

# 1. Run a cycle
swarma cycle starter --topic "test run"
# Expected: table showing agent outputs + costs

# 2. Check status
swarma status
# Expected: teams listed, recent run shown, costs displayed

# 3. Check a real squad (if installed)
swarma team show hook-lab
# Expected: team config with agents, flow, metrics

If all three pass, the GROWS loop is operational.

What swarma Is Not