Local MCP server for the LLM Council workflow. Exposes tools to run a 3-stage LLM Council workflow (multi-model decision-making) directly against OpenRouter without Docker or network dependencies beyond reaching OpenRouter.
Two tools with well-structured schemas and detailed descriptions. Both tools have proper Pydantic Field annotations with min/max length constraints and clear parameter documentation. Descriptions are substantive (150-400 chars) and explain WHAT the tool does and WHEN to use it. However, output schemas are not formally documented, responses are free-form strings rather than structured objects. Error handling is present but generic ('Error: ...'). Tool names are clear action verbs (ask_*). No critical security issues detected; API keys are handled server-side via OpenRouter integration.
Get a multi-model 'second opinion' on an important decision. Runs the full 3-stage council process locally and synchronously: 1. Each council model (3 by default) answers the question independently, in parallel. 2. Each model anonymously reviews and ranks the others' answers (so a model can't recognize or favor its own response). 3. A chairman model reads everything from stages 1 and 2 and writes one final, synthesized answer. If the chairman call fails, the code automatically falls back to trying other council models as chairman until one succeeds. Returns a Markdown report with the chairman's final synthesized answer, followed by each individual council member's Stage-1 response for transparency.
Get a multi-perspective single-model 'second opinion' on a decision without OpenRouter API keys. Generates a structured prompt instructing you (the calling LLM) to simulate 5 different viewpoints (Pragmatist, Architect, Skeptic, Clean Code Advocate, UX Thinker), debate them internally, and synthesize a final balanced decision.
Output schemas not documented. Both tools return free-form strings (Markdown reports) with no structured schema definition. LLMs cannot reliably extract fields or plan downstream operations. Should document return type as structured object with fields like 'chairman', 'verdict', 'council_responses', 'stage_details'.
Error responses are generic strings ('Error: ...') without recovery guidance. LLMs cannot determine if errors are retryable, user-fixable, or fatal. Should categorize errors and suggest next steps (e.g., 'API key missing. Set OPENROUTER_API_KEY environment variable.').
No pagination or result limits documented. ask_council returns full Markdown report with all council responses. For large deliberations, this could exceed context windows. Should cap response length and offer pagination or summary modes.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
ask_council accepts optional 'models' and 'chairman' parameters but does not validate them against available OpenRouter model IDs. Invalid model names will fail at runtime with opaque API errors. Should validate model IDs upfront and return actionable error messages.
ask_internal_council returns a prompt string for the calling LLM to execute, but does not document the expected format or structure of that prompt. Callers cannot reliably parse or validate the output. Should document prompt structure and consider returning structured metadata alongside the prompt.