polyphony — Claude Code skill: what it does, how to install
polyphony gives every AI agent a separate Docker container and git branch, so parallel agents never step on each other's files.
MIT 116 installs ★ 713 updated 2026-10-13
npx skills add alinaqi/maggy --skill polyphony -a claude-codeWhat polyphony does
polyphony gives each agent a separate Docker container holding a full git clone at /workspace and a branch of its own. Agents can't collide, tests run independently, and every branch yields a clean PR. The skill loads automatically when container isolation is available, and /spawn-team uses it by default.
Tasks come from GitHub Issues or a local SQLite queue. A router scores complexity from 0 to 10 across five dimensions, including security boundaries and concurrency. Scores 0 to 3 go to Kimi alone, 4 to 6 add a Codex review, and 7 to 10 go straight to Claude. Rules in routing.yaml are checked top-down and the first match wins.
Settings live in ~/.polyphony/, and polyphony init creates the templates. Without a running Docker or OrbStack there are no containers, so nothing is isolated. You also need at least one agent CLI (Claude, Codex or Kimi) configured with a subscription rather than an API key.
What it helps with
- keep parallel agents on separate branches with no merge conflicts
- run each agent's tests independently in its own container
- send a simple task to Kimi and a risky one to Claude based on a complexity score
- pull tasks from GitHub Issues into an agent queue
- get a clean PR out of every agent session
Example prompts for the agent
- Spin up a team with
/spawn-team, each agent in its own Docker container - Run
polyphony spawn "Fix auth bug"and then check progress withpolyphony status - Add a rule to
routing.yamlthat sends risk: high tasks to claude
How to install polyphony
1Skills CLI (Vercel)
npx skills add alinaqi/maggy --skill polyphony -a claude-code2Manually: put the files in place
~/.claude/skills/polyphony/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/polyphony/
1Skills CLI (Vercel)
npx skills add alinaqi/maggy --skill polyphony -a codex2Manually: put the files in place
.agents/skills/polyphony/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add alinaqi/maggy --skill polyphony -a cursor2Manually: put the files in place
.cursor/skills/polyphony/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install alinaqi/maggy skills/polyphony2Manually: put the files in place
.github/skills/polyphony/SKILL.md1Skills CLI (Vercel)
npx skills add alinaqi/maggy --skill polyphony -a gemini-cli2Platform CLI
gemini skills install https://github.com/alinaqi/maggy --path skills/polyphonyManually: put the files in place
.devin/skills/polyphony/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add alinaqi/maggy --skill polyphony -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about polyphony
How do I install polyphony in Claude Code?
Run npx skills add alinaqi/maggy --skill polyphony -a claude-code. The skill file ends up at ~/.claude/skills/polyphony/SKILL.md.
What does polyphony do?
It runs each agent in a separate Docker container with a full clone of the repo and its own branch. A router decides whether Kimi, Codex or Claude takes the task.
What does polyphony need to work?
A running Docker or OrbStack and at least one agent CLI (Claude, Codex or Kimi) configured with a subscription. polyphony init creates the config templates in ~/.polyphony/.
Source text
SKILL.md · MIT
---
name: polyphony
description: Multi-agent orchestration with container-isolated workspaces — each agent session runs in its own Docker container with independent git branches
when-to-use: Always loaded when container isolation is available (Docker/OrbStack installed). Default for /spawn-team.
user-invocable: false
effort: high
---
# Polyphony — Multi-Agent Orchestration
Container-isolated workspaces for parallel agent execution. Each agent gets its own Docker container with a full git clone on its own branch. No conflicts, independent tests, clean PRs.
---
## Architecture (6 Layers)
1. **Work Source** — Tasks from GitHub Issues (`gh api`) or local SQLite queue
2. **Orchestrator** — Supervisor loop: discover -> claim -> route -> provision -> run -> verify -> land
3. **Router** — Pure function: Task x Policy -> RunSpec (5-dimension complexity scoring)
4. **Identity Broker** — Resolves named credentials to volume mounts + env overlays
5. **Workspace Manager** — Per-task `git clone --reference`, branch checkout, cleanup
6. **Worker Runtime** — Docker container create/start/stop/logs lifecycle
---
## Task Lifecycle
```
DISCOVERED -> CLAIMED -> ROUTED -> PROVISIONED -> RUNNING -> VERIFYING -> LANDED
| |
v v
FAILED --> BLOCKED
|
v
CLAIMED (retry)
```
---
## Prerequisites
- Docker or OrbStack installed and running
- At least one agent CLI available (Claude, Codex, or Kimi)
- CLI subscriptions configured (not API keys)
Check:
```bash
command -v docker &>/dev/null || command -v orbctl &>/dev/null
```
---
## Configuration
All config lives in `~/.polyphony/`:
| File | Purpose |
|------|---------|
| `config.yaml` | Workspace root, poll interval, max concurrency |
| `identities.yaml` | Named credential bundles with volume paths |
| `agents.yaml` | Agent profiles (CLI commands, strengths) |
| `routing.yaml` | Routing rules and fallback chains |
Initialize with: `polyphony init`
---
## Routing Rules
Rules are evaluated top-down; first match wins. Each rule has a `match` predicate and an `agent` target.
```yaml
rules:
- match: { task_type: docs, risk: low }
agent: kimi
- match: { task_type: bugfix }
agent: codex
- match: { risk: high }
agent: claude
default:
agent: claude
fallback: [codex, kimi]
```
---
## Complexity Scoring (5 Dimensions)
Each dimension scores 0-2. Total 0-10.
| Dimension | Source |
|-----------|--------|
| Cyclomatic depth | LOC + scope size |
| Fan-out | Number of callers |
| Security boundary | Auth/PII keywords |
| Concurrency | Lock/transaction keywords |
| Domain invariants | Risk level + task type |
Routing thresholds:
- **0-3**: Delegate to Kimi solo
- **4-6**: Kimi + Codex review
- **7-10**: Claude direct
---
## Container Isolation
Each task gets:
- Its own Docker container from `polyphony-worker:latest`
- A full git clone at `/workspace` (not a worktree)
- Auth volumes mounted read-only (e.g., `~/.claude:/home/worker/.claude:ro`)
- Independent test execution
- Its own branch for PRs
---
## CLI Commands
```bash
polyphony init # Create ~/.polyphony/ with config templates
polyphony spawn "Fix auth bug" # Create and route a task
polyphony status # Show task states
polyphony cleanup # Remove completed workspaces
```
---
## Integration with Existing Skills
- **cross-agent-delegation**: Uses Polyphony's complexity scoring for routing decisions
- **agent-teams**: Uses Polyphony's workspace isolation instead of shared directories
- **spawn-team**: Uses Polyphony's container provisioning for feature agents
© alinaqi · MIT · alinaqi/maggy