Bernstein

What sipyourdrink-ltd/bernstein MCP server does

The sipyourdrink-ltd/bernstein MCP server makes Bernstein's workflow orchestrator available to MCP clients. Bernstein coordinates multiple command-line agents from a declarative YAML workflow. A workflow can define phases, roles, dependencies, and conditions that control when each node runs. The resulting work can be source-code changes, research reports, datasets, operational results, or audit evidence.

The orchestration layer is intended for local execution rather than a hosted data plane. It supports more than 40 CLI agent adapters, including Claude Code, Codex, Cursor, Aider, Gemini CLI, GitHub Copilot CLI, OpenHands, OpenCode, Goose, Qwen, and Ollama. A generic prompt wrapper is also available according to the project description.

How it works

Bernstein evaluates the workflow graph with a deterministic scheduler implemented in Python. Models do not choose the next task or alter the coordination plan, so the same workflow can be replayed and compared with the original run. Independent nodes can run in parallel, while dependencies and phase rules control ordering and access.

Coding tasks receive separate Git worktrees by default, preventing agents from sharing mutable checkout state. Non-code tasks use workspaces under .sdd/workspaces/. Tasks claim backlog entries atomically, and merge gates can control how coding changes are integrated. The project also describes optional sandbox backends for stricter filesystem controls.

Execution records include a replay journal and lineage information. Enabling BERNSTEIN_AUDIT=1 adds an HMAC-chained audit log whose receipts can be verified offline. Artifact-mode tasks can finish with a signed lineage receipt instead of a Git commit, allowing the same governance model to cover reports, datasets, action logs, and other deliverables.

Setup and configuration

The sipyourdrink-ltd/bernstein MCP server is distributed as an open-source Python project targeting Python 3.12 or newer. The README identifies both a PyPI package and a container image, and documents MCP server operation over stdio and HTTP/SSE. The supplied material does not include the exact launch command or a complete MCP client configuration block, so installation commands should be taken from the project's installation documentation rather than inferred here.

Configuration is centered on YAML workflow files. These files specify the run graph, phases, roles, dependencies, and artifact expectations. The audit log is opt-in through the BERNSTEIN_AUDIT environment variable. Bernstein also includes an air-gap installation profile for environments that cannot use an external network path.

Limitations and notes

The project labels itself beta and says it is solo-maintained and under active development. Minor releases may change interfaces, so the README recommends pinning versions for dependent deployments. Worktree isolation can be disabled, in which case tasks use a shared checkout. The provided material does not establish support for a particular named MCP desktop client, nor does it describe authentication requirements for the MCP endpoints.