The same multi-model pipeline behind the web app, as a local command-line tool — and as native MCP tools your coding agent can call. Local-first, your own API keys, and a free offline mode so you can prove the wiring before spending a cent.
Requires Python 3.11+ and git.
git clone https://github.com/adta7/ai-orchestrator.git && cd ai-orchestrator
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[providers]"
The fake council runs the whole pipeline with deterministic stand-in models,
so you can confirm everything works for zero cost.
orchestrate relay "design a URL shortener" --council fake
orchestrate ask "Monolith or microservices for our MVP?" orchestrate brainstorm "names for a developer-tools startup" --fanout 8 orchestrate rhetoric "Should we rewrite the monolith?" --agents 3 --rounds 4 orchestrate relay "design a rate limiter" --council orchestra orchestrate review --artifact-file plan.md --council cli --json
Add --council lite for a faster, cheaper run, or --json
for machine-readable output. Full transcripts are saved locally under the project.
Install with the mcp extra, then register the server so Claude Code (or any
MCP-capable agent) can call the council as native tools.
pip install -e ".[providers,mcp]"
claude mcp add -s user -t stdio orchestrator -- "$(pwd)/.venv/bin/orchestrate-mcp"
Design or plan something — research → architecture → plan → independent review → synthesis. Returns a plan; never edits files.
Independently critique a draft or plan with multiple models. Returns a verdict + prioritized fixes.
Answer a high-stakes question by synthesizing several models, preserving disagreement.
The fake council is free; for real models, add the keys you have to a local
.env (git-ignored) in the repo:
ANTHROPIC_API_KEY=... OPENAI_API_KEY=... GOOGLE_API_KEY=...
Your keys stay on your machine — prompts go only to the model providers you configure. Don't paste keys into chat.