Labsco
MCP SERVER

Collusion

by nideweilaiya

Three agents design the same thing from different angles, review each other, and hand back a ranked top three instead of one answer.

Reasoning Scaffolds & Agent Workflow Engines
Summary
The value is in the disagreement, not the volume.

A single model asked for a design tends to produce one confident answer with no visible alternatives, and no pressure to justify scope. Running three roles against each other creates that pressure — the feasibility pass in particular exists to cut architectures that ignore a stated constraint. Two limits to know before adopting it: the LLM adapter is DeepSeek-only today, and the three roles are fixed, so custom perspectives are not yet available.

What it is

A multi-agent design engine exposed over MCP. Give it a task and it decomposes the requirement, fills in the parts you did not mention, has three role-based agents each produce a full proposal, makes them cross-review one another, forces a feasibility pass that cuts over-engineering, then scores everything on five dimensions and returns the top three.

What you get
  • `brainstorm_orchestrate` runs the whole pipeline from a task description; `agents` controls how many perspectives run, from a single fast pass up to the full three
  • Because a full run takes minutes, it is asynchronous: `brainstorm_status` reports progress by task ID and `brainstorm_result` returns the finished output
  • Gap detection — the requirement list is audited before any design starts, and missing pieces such as security, deployment or migration are added
  • A feasibility brake: an engineering role cuts proposals that exceed the stated constraints, and an over-built design is pushed down the ranking rather than quietly returned
  • Scoring on correctness, completeness, feasibility, innovation and business alignment, with the reasoning attached to each score
  • Output as a Markdown blueprint plus a self-contained HTML report; `collusion_blackboard_status` reports on shared-blackboard runs
Requirements

Python 3.10 or higher and a DeepSeek API key. The README's install line names `collusion-mcp`, which PyPI does not have; that is also the console script the project declares, at version 0.5.1. The key is read from `DEEPSEEK_API_KEY`, falling back to `OPENAI_API_KEY` then `LLM_API_KEY`, or it can go in a `config.json`. Clients launch the server module over stdio; an SSE mode is started with `--sse --port 8020`.