Turn one AI assistant into a coordinated dev team. Works with Cursor, Claude Code, Copilot, Windsurf, and any MCP client.
npx -y github:cloga/optimus-code init
Copied!
A self-evolving agent lifecycle
Your AI assistant becomes a Master Agent that analyzes tasks and selects the right specialists.
Tasks are dispatched to specialized roles — Architect, PM, QA, Dev — that work in parallel via the Council pattern.
Agents start ephemeral (T3), precipitate into templates (T2), and freeze as reusable instances (T1). Your team grows with every task.
From request to merge — a fully automated pipeline
You describe what you need in natural language
Product Manager structures requirements and creates a tracking issue
Developer implements with full architectural context
QA Engineer runs tests and reviews quality
Pull request merged, issue closed, work shipped
Real workflows, not demos
You debug alone, context scattered across files and tabs.
PM creates issue, Architect designs fix, Dev implements, QA verifies — all tracked.
Solo agent guesses at architecture and writes everything in one pass.
Council debates architecture, then Dev builds with full context and reviewed design.
One reviewer, one perspective, blind spots everywhere.
3 expert reviewers in parallel with synthesis report — security, performance, and architecture.
Optimus Code is built on a Great Unification architecture — a pure Node.js MCP daemon with self-evolving agents, map-reduce councils, and issue-first SDLC. Read the full technical deep-dive.
Orchestrators plan, developers code. Separation of concerns enforced by the system.
Spawn parallel expert reviews. Map-reduce consensus from multiple perspectives.
Roles define WHO. Skills define HOW. Many-to-many binding at runtime.
Parent-child traceability across all GitHub Issues. Every delegation tracked.
Continuous learning from project history. Past mistakes never repeated.
Ephemeral → Template → Instance. Your team grows with every task.
How our multi-agent architecture compares to Claude Code's new Agent Teams
| Dimension | Claude Code Agent Teams | Optimus Swarm |
|---|---|---|
| Multi-Vendor | — Claude only (single vendor lock-in) | ✓ Claude + GPT + Gemini + Qwen + any ACP agent |
| Cross-IDE | — Terminal only (requires tmux/iTerm2) | ✓ VS Code, Cursor, Windsurf, Claude Code, Codex, any MCP client |
| Expert Review | — Same-model Lead reviews same-model Worker | ✓ Cross-model heterogeneous council (Map-Reduce + auto VERDICT) |
| Design Process | — Worker plans → Lead approves (same model) | ✓ Problem-First SDLC: 3+ experts propose → Master synthesizes → Council reviews |
| Coordination | — Shared task list, workers self-claim | ✓ PM-orchestrated: intelligent role-aware assignment with full context |
| Communication | — Peer-to-peer messaging between agents | ✓ Centralized orchestration — eliminates cross-model protocol issues |
| Nested Delegation | × Not supported (flat topology) | ✓ Up to 3 levels of nested delegation with depth tracking |
| Persistent Memory | × None | ✓ Project + Role + User level memory with relevance scoring |
| SDLC Integration | — Task execution only | ✓ Full lifecycle: Problem → Proposal → Solution → Dev → Review → Release |
| GitHub/ADO Integration | × None | ✓ Auto Issue/PR creation, lineage tracking, labels |
VS Code users are auto-configured. Others: add to your MCP client config.
Any MCP-compatible AI coding tool. No lock-in.