An MCP server that runs a 3-stage LLM council process to gather diverse perspectives, perform peer ranking, and synthesize a final answer using a chairman model
Single tool 'consult_council' has a reasonable description and input schema structure, but critical gaps in parameter documentation, output schema definition, and error handling guidance. The schema shows input parameters with types and descriptions, but lacks comprehensive validation documentation, output structure documentation, and error recovery patterns. Parameter descriptions exist but are minimal. No evidence of idempotency guarantees, retry guidance, or actionable error messages. The tool's description is adequate (195 chars) but lacks WHEN/WHY guidance and doesn't address the asynchronous nature of the council process or expected response structure. File attachment parameters ('files' array) lack validation constraints (max count, max size, allowed formats).
Ask the LLM Council for a fresh perspective on a problem. The Council will run a 3-Stage Process: 1. Exploration (Multiple Models) 2. Peer Review (Ranking) 3. Synthesis (Chairman's Verdict)
Output schema not documented. Tool description does not specify what fields are returned (verdict, rankings, model responses, confidence scores, etc.) or their structure. LLMs cannot plan downstream actions or extract specific data.
Parameter 'files' lacks validation constraints. No maximum count, file size limits, or allowed formats specified. Description is vague ('List of file contents (strings)'). LLMs cannot validate input before calling, risking failed requests or resource exhaustion.
No error recovery guidance. Tool description does not explain what to do if council fails to reach consensus, a model times out, or insufficient context is provided. Error messages will be opaque to the LLM.
Missing idempotency and retry documentation. Given the 3-stage async process and network dependencies, the tool should declare whether repeated calls with identical inputs produce the same result and whether it's safe to retry on timeout.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 56 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 33 | - | v1 |
Tool description lacks action verb specificity. 'Ask the LLM Council' is vague about what happens, does it block until completion? Return immediately with a task ID? The agent cannot plan correctly without knowing the interaction model.
Parameter 'query' description is minimal ('The question or problem description'). Should specify format constraints (min/max length), what constitutes a valid question, and whether multi-turn context is supported.
No output field chaining documented. If downstream tools need model names, ranking scores, or disagreement metrics, the output must include them. Currently unknown what fields exist.