# sipyourdrink-ltd/bernstein [Health: Active]

**Category:** 🤖 Coding Agents  
**Repository:** https://github.com/sipyourdrink-ltd/bernstein  
**GitHub Stars:** 1326  
**Views:** 6  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/sipyourdrink-ltd-bernstein

## Description
Deterministic multi-agent orchestrator for 37 CLI coding agents (Claude Code, Codex, Cursor, Aider, Gemini CLI, GitHub Copilot CLI, OpenHands, OpenCode, Goose, Qwen, Ollama, and more). MCP server mode (stdio + HTTP/SSE) exposes the orchestrator to any MCP client. Git worktree isolation per agent, HMAC-chained audit trail, cost-aware model routing via contextual bandit. Apache 2.0.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "bernstein": {
    "command": "uvx",
    "args": ["bernstein"]
  }
}
```

## Documentation

## What sipyourdrink-ltd/bernstein does

sipyourdrink-ltd/bernstein provides an MCP-facing governance layer for coordinating multiple AI agents. Its scheduler is implemented in Python rather than delegated to a language model, so the task graph and execution decisions can be reproduced from the workflow definition.

A workflow is described in YAML with phases, roles, dependencies, and execution conditions. The same framework can coordinate coding tasks and other deliverables, including research reports, datasets, action logs, operational results, and audit evidence packs. Agents can be selected from a broad set of CLI adapters, including Claude Code, Codex, Cursor, Gemini CLI, Aider, GitHub Copilot CLI, OpenHands, OpenCode, Goose, Qwen, and Ollama.

## How it works

An MCP client connects to Bernstein through either stdio or HTTP/SSE. Bernstein then runs the declared workflow locally, starting eligible tasks according to the dependency graph and policy rules. Scheduling does not depend on an LLM choosing the next step.

Coding tasks receive separate Git worktrees by default, while artifact-oriented tasks use workspaces beneath `.sdd/workspaces/`. Tasks claim backlog state atomically, so agents do not share a mutable working directory unless worktree isolation is disabled. Merge gates and role restrictions can control what proceeds between phases.

Execution history is captured in a replay journal and a lineage record. The optional HMAC-chained audit log can add verifiable receipts when `BERNSTEIN_AUDIT=1` is enabled. Hash mismatches identify the step where a replay diverges, allowing records to be checked offline from the generated artifacts.

## Setup and configuration

Bernstein requires Python 3.12 or newer. The project publishes a Python package and a container image, and also provides an air-gap installation profile. Configure the MCP client to launch the server over stdio or connect to its HTTP/SSE endpoint, depending on the deployment layout.

Workflow definitions are YAML files. They specify the run name, phases, permitted roles, dependencies, and conditions for execution. Worktree isolation is enabled by default for coding tasks; stricter filesystem controls are available through optional sandbox backends. Set `BERNSTEIN_AUDIT=1` when the opt-in HMAC audit trail is needed.

## Tools and capabilities

- Coordinate multiple CLI coding agents from one workflow.
- Run independent tasks in parallel using a deterministic scheduler.
- Apply role, phase, dependency, approval, and merge gates.
- Isolate coding work with per-task Git worktrees.
- Produce replayable journals, lineage records, and optional HMAC audit receipts.
- Route work across more than 50 selectable CLI agent adapters and a generic prompt wrapper.
- Handle non-code artifact contracts such as reports, datasets, and audit evidence packs.

## Limitations and notes

The project is marked beta and is maintained by a solo maintainer. Minor releases may change interfaces, so deployments that depend on Bernstein should pin a version. The scheduler and audit records are local and file-based; the project does not use a SaaS data plane.

Hardware-key evidence features include offline PIV custody and FIDO2 approval verifiers, with pinned trust and explicit time. Token custody does not increase TRACE trust levels. Worktree isolation can be disabled, but doing so makes tasks run in the shared checkout rather than separate worktrees.

_Full upstream README: https://allmcps.com/mcp/sipyourdrink-ltd-bernstein/readme_