Repository URL: https://github.com/sipyourdrink-ltd/bernstein
Category: Developer Tools — orchestration / coding agents
Brief description:
Bernstein is an orchestrator for CLI coding agents that exposes its control surface over MCP, so an MCP client can plan, spawn, supervise, and audit multi-agent runs as tool calls.
- Install:
pip install bernstein
- Run (stdio):
python -m bernstein.mcp
- Remote transport: also ships an HTTP/streaming transport with OAuth, for clients that don't spawn a local subprocess.
- License: Apache-2.0
- Language: Python (>=3.12)
Tools exposed (25). Run control (bernstein_run, bernstein_run_status, bernstein_stop, bernstein_cancel), task lifecycle (bernstein_tasks, bernstein_claim, bernstein_update, bernstein_complete, bernstein_create_subtask, bernstein_task_handle, bernstein_task_capsule), human-in-the-loop approval (bernstein_approve), scenarios (bernstein_scenarios, bernstein_scenario, bernstein_scenario_status), messaging and artifacts (bernstein_post_message, bernstein_post_artifact), operations (bernstein_health, bernstein_status, bernstein_cost, bernstein_context, bernstein_shutdown_orchestrator), skills (load_skill), and verification (bernstein_verify_lineage, verify_chain).
What distinguishes it from a generic task-runner MCP server: scheduling is deterministic — there is no model in the coordination loop, so the same inputs produce the same task graph. Each agent runs in its own git worktree, and every step is recorded in an HMAC-signed audit chain. The last two tools above let a client verify that chain directly, so the run's history can be checked rather than taken on trust. A lineage resource is also exposed for reading those records.
It wraps existing CLI coding agents rather than replacing them — Claude Code, Codex, Gemini CLI, Aider, Qwen, OpenHands, Goose, Cline and others are supported through adapters, so the MCP tools drive whichever agent is configured.
Docs: https://github.com/sipyourdrink-ltd/bernstein/tree/main/docs
Repository URL: https://github.com/sipyourdrink-ltd/bernstein
Category: Developer Tools — orchestration / coding agents
Brief description:
Bernstein is an orchestrator for CLI coding agents that exposes its control surface over MCP, so an MCP client can plan, spawn, supervise, and audit multi-agent runs as tool calls.
pip install bernsteinpython -m bernstein.mcpTools exposed (25). Run control (
bernstein_run,bernstein_run_status,bernstein_stop,bernstein_cancel), task lifecycle (bernstein_tasks,bernstein_claim,bernstein_update,bernstein_complete,bernstein_create_subtask,bernstein_task_handle,bernstein_task_capsule), human-in-the-loop approval (bernstein_approve), scenarios (bernstein_scenarios,bernstein_scenario,bernstein_scenario_status), messaging and artifacts (bernstein_post_message,bernstein_post_artifact), operations (bernstein_health,bernstein_status,bernstein_cost,bernstein_context,bernstein_shutdown_orchestrator), skills (load_skill), and verification (bernstein_verify_lineage,verify_chain).What distinguishes it from a generic task-runner MCP server: scheduling is deterministic — there is no model in the coordination loop, so the same inputs produce the same task graph. Each agent runs in its own git worktree, and every step is recorded in an HMAC-signed audit chain. The last two tools above let a client verify that chain directly, so the run's history can be checked rather than taken on trust. A lineage resource is also exposed for reading those records.
It wraps existing CLI coding agents rather than replacing them — Claude Code, Codex, Gemini CLI, Aider, Qwen, OpenHands, Goose, Cline and others are supported through adapters, so the MCP tools drive whichever agent is configured.
Docs: https://github.com/sipyourdrink-ltd/bernstein/tree/main/docs