Bernstein is an open-source multi-agent orchestration system that coordinates AI coding agents (like Claude Code, Codex, Gemini CLI) to work in parallel on your codebase. It decomposes goals into tasks, assigns them to agents with isolated git worktrees, and verifies results through quality gates before merging.

How do I install Bernstein?

Install Bernstein with pipx: 'pipx install bernstein'. You can also use pip or uv. Requires Python 3.12+ and at least one supported CLI coding agent.

What AI coding agents does Bernstein support?

Bernstein supports 29 agents including Claude Code, Codex CLI, Gemini CLI, Cursor, Aider, Amp, Roo Code, Kiro, Qwen, Goose, Ollama, and more. It also has a generic adapter for wrapping any CLI tool.

How does Bernstein differ from CrewAI or AutoGen?

Bernstein's orchestrator is deterministic Python code — no LLM tokens are spent on coordination. It works with real CLI coding agents, provides git worktree isolation, runs quality gates on every result, and includes cost tracking. CrewAI and AutoGen use LLM-driven coordination and don't support CLI agents.

Yes. Bernstein is open-source under the Apache 2.0 license. You only pay for the AI model API usage of the agents themselves.

Orchestrate parallel AI agents on your codebase

Run Claude Code, Codex, and Gemini CLI simultaneously. Deterministic scheduling, quality gates, cost tracking. You come back to working code.

View on GitHub Documentation

$ pipx install bernstein

110+

GitHub stars

7.7k

Monthly downloads

Agent adapters

2,600+

Tests passing

How it works

Three steps. No babysitting.

Set a goal

Describe what you want built. Bernstein decomposes it into tasks with the right roles, models, and dependencies.

Agents work in parallel

Each agent gets its own git worktree. Opus for architecture, Sonnet for implementation, Haiku for tests. Automatically routed.

Verified and merged

Quality gates run lint, types, and tests on every result. Only verified work gets merged. Failed tasks retry with escalated models.

Works with every major coding agent

Mix local models for boilerplate with cloud models for architecture. In the same run.

Claude Code

Opus, Sonnet, Haiku

Codex CLI

GPT-5.4, GPT-5.4-mini

Gemini CLI

Gemini 3.1 Pro, 3 Flash

Cursor

Any model via Cursor

Aider

Any OpenAI/Anthropic

Ollama

Local, fully offline

Amp

Sourcegraph Amp

Goose

Block Goose CLI

Roo Code

VS Code extension CLI

Kiro

AWS Kiro CLI

Qwen

Alibaba Qwen Agent

+18 more

Generic CLI adapter

Built for production use

Not a demo. A system you can run unsupervised.

Deterministic scheduling

The orchestrator is pure Python. Zero LLM tokens on coordination. Predictable, debuggable, fast.

29 agent adapters

Claude Code, Codex, Gemini CLI, Aider, and 25 more. Mix models in one run. Switch providers without changing config.

Quality gates

Lint, type check, tests, security scan, architecture conformance. All run before merge. Failed work retries automatically.

Cost-aware routing

Epsilon-greedy bandit learns which model works best per task type. Typical savings of 50-60% vs uniform model selection.

Git worktree isolation

Each agent works in its own worktree. No merge conflicts between agents. Clean, linear commit history.

Full observability

Per-agent cost tracking, token monitoring, quality trends, Prometheus metrics. Know what happened and what it cost.

How it compares

Different category, different architecture.

	Bernstein	CrewAI	AutoGen	LangGraph
Orchestrator	Deterministic code	LLM-driven	LLM-driven	Graph + LLM
CLI agent support	29 adapters	No	No	No
Git isolation	Worktrees	No	No	No
Quality gates	Built-in	No	No	Partial
Cost tracking	Per-agent	No	No	No
Self-evolution	Built-in	No	No	No