Autonomous AI engineering command center.

Chat with your codebases across Claude, Codex, and Kimi — with semantic routing that picks the cheapest model that can handle the task.

• Semantic Routing — local Ollama model rates task complexity and type, routes to Local/Kimi/Codex/Claude accordingly
• Interactive Chat — SSE streaming, session resume, ghost-text suggestions (Tab to accept)
• Inline Model Forcing — type "use claude" in any message to override routing
• Parallel Execution — Polyphony container orchestration for complex tasks (blast>=7)
• AI-prioritized Tasks — ranks open issues by urgency + OKR alignment
• One-click Execute — TDD pipeline with iCPG context enrichment
• Council PR Review — a multi-model council reviews a GitHub PR from the dashboard: deterministic mega-PR chunking, a static gate (tsc/ruff) as ground truth, an adversarial refute pass to kill false positives, and extensible per-language skills (Python/TS/Go/Rust/Java/C#/Ruby/PHP — drop in more). pip install maggy-harness[review]
• Competitor Intelligence — auto-discovered competitors, daily AI briefing
• Engram Memory — persistent cross-session memory with amnesia diagnostics

pip install maggy-harness    # or: pipx install maggy-harness
maggy bootstrap              # installs skills, hooks, ~/bin model wrappers, plugins

# Pull the local model (optional but recommended)
ollama pull qwen3-coder:30b-a3b

# Configure (optional — Maggy runs in local mode without keys)
export GITHUB_TOKEN=ghp_...
export ANTHROPIC_API_KEY=sk-ant-...

# Launch
maggy              # interactive REPL (auto-detects project)
maggy serve        # web dashboard at localhost:8080
maggy chat api     # chat with a specific project

Or install from source for development:

cd maggy
pip install -e .

Or from inside Claude Code:

/maggy-init   # interactive setup wizard
/maggy        # launch dashboard

Every message gets a semantic blast score (1-10), then routes to the cheapest capable model:

The router learns from outcomes — successful tasks reinforce the routing decision.

Run multiple agents concurrently without conflicts. Each agent session runs in its own Docker/OrbStack container with a full git clone on its own branch — six layers: work source → orchestrator → router → identity broker → workspace manager → worker runtime.

• Auto-isolation — a second Claude Code session in the same project automatically provisions its own workspace (no setup).
• Container mode is the default when Docker + the polyphony CLI are present; falls back to native parallel agents otherwise.

polyphony init                 # one-time: create ~/.polyphony/ config
polyphony spawn "add auth"     # create + route a task to an agent
polyphony status               # running agents / task states
polyphony cleanup              # remove completed workspaces

From Claude Code: /polyphony-init, /polyphony-spawn, /polyphony-status, or /spawn-team (a coordinated TDD team, container-isolated by default). Requires Docker or OrbStack. Full spec: Polyphony spec.

cd maggy
python3 -m pytest tests/ -x -q

887 tests, target coverage >= 80%.

• Full CLI & REPL reference
• Architecture (v5)
• Polyphony spec
• Mnemos implementation

MIT