---
id: bernstein
version: "3.15.1"
license: Apache-2.0
license_treatment: permissive
maintenance: active
---

# bernstein — Deterministic orchestrator for CLI coding agents (Claude Code, Codex, Gemini CLI, and 40+ more). No model in the coordination loop, so parallel runs in per-task git worktrees replay byte-identically. Signed lineage plus an opt-in HMAC audit chain a reviewer checks offline, without rerunning it. Cluster mode, air-gap deploy.

License: permissive · Maintenance: active · Downloads: 199.6K/mo

## What it is and what it does

Bernstein is a deterministic orchestrator for running multiple CLI coding agents in parallel. It coordinates tasks without an LLM in the scheduling loop—meaning a run is reproducible byte-for-byte when replayed with the same inputs and agent outputs. Each coding task runs in its own git worktree, isolated by default; coordination state (the task backlog) is shared and claimed atomically. The package records every run in a replay journal and maintains an always-on lineage spine that tracks lineage-bearing steps; an optional HMAC-chained audit log adds cryptographic receipts you can verify offline without rerunning the orchestration.

It ships with adapters for Claude Code, Codex, Gemini CLI, and 40+ other agents, plus a generic `--prompt` wrapper. You can watch runs live via a terminal dashboard (`bernstein live`) or browser dashboard (`bernstein gui serve`), and verify determinism by replaying a recorded run and checking for divergence. The package includes air-gap deployment support and is built on fastapi, cryptography, and opentelemetry for observability. It's beta software, solo-maintained, and actively developed; minor versions may change interfaces.

Use it for:

- Run multiple coding agents on the same task in parallel and compare their outputs deterministically, with cryptographic proof of what each agent produced.
- Verify that a multi-agent workflow produces identical results when replayed with the same inputs, catching non-determinism at the exact step it occurs.
- Build compliance-auditable agent workflows where every step is signed and can be verified offline by a reviewer without access to live infrastructure.
- Coordinate agent tasks across a cluster with per-task isolation, ensuring agents cannot interfere with each other's working directories or shared state.
- Record and replay agent-driven code generation or testing workflows to debug failures or reproduce issues from a previous run.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Bernstein orchestrates parallel CLI coding agents (Claude Code, Codex, Gemini CLI, and 40+ others) with deterministic replay, cryptographic lineage tracking, and optional audit chains—no LLM in the coordination loop, so runs are reproducible end-to-end.

Yes, if you are orchestrating multiple CLI coding agents and need deterministic replay, cryptographic auditability, and per-task isolation. The low install friction and active maintenance support the use case. No, if you need a stable API—beta status means minor versions may break interfaces, so pin carefully. No, if you are not using CLI agents or do not need replay and audit guarantees.

## Install

pip install bernstein
uv add bernstein
poetry add bernstein

## Installing bernstein

Before you install: Low friction: pure Python wheel with 34 runtime dependencies including fastapi, cryptography, and opentelemetry. Active maintenance (last commit 2026-08-14, 882 stars); beta status means minor versions may change interfaces—pin the version if you depend on it. Requires Python 3.12+.

License in practice: Apache-2.0 permissive license; you can use, modify, and distribute freely in commercial and private projects, with attribution and no warranty.

Quickstart:

# Install
uv tool install bernstein  # or: pipx install bernstein

# Initialize and verify
bernstein init
bernstein doctor

# Run an agent task
bernstein -g "fix the failing test in tests/test_foo.py"

# Verify determinism
bernstein replay latest --verify

Requires Python 3.12 or later. Requires at least one CLI agent installed and authenticated (e.g., Claude Code, Codex, Gemini CLI). Git must be available for per-task worktree isolation.

Verify before relying:

- Whether all 40+ agent adapters are documented or discoverable in the package itself.
- Performance characteristics when running many agents in parallel (latency, resource overhead).
- Compatibility of the air-gap install profile with all runtime dependencies.
- Whether the HMAC audit chain and signed receipts are suitable for regulated compliance workflows.

## Package facts

- License: Apache-2.0 (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 199.6K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

multi-agent orchestration cli, deterministic agent replay, coding agent coordination, audit trail provenance tracking, parallel llm agent runner, reproducible workflow orchestration, agent task scheduling, agent-orchestration, deterministic-replay, audit-trail

[View on SkillFed](https://skillfed.io/packages/bernstein) · [View on PyPI](https://pypi.org/project/bernstein/)