Descartes — doubt, until what remains is certain. An MCP server that doubts every plan decision until it converges, grounded in real evidence.
Three tools with reasonable descriptions but critical schema and parameter documentation gaps. doubt() and ground() lack explicit input schema visibility in the provided code; verdict() has a vague object parameter with no field documentation. Descriptions are lengthy (100-200+ chars) but lack actionable constraints, enums, or format guidance. No error handling guidance, no output schema documentation, and no parameter type constraints visible. Tool names are verb-based (doubt, ground, verdict) but lack clarity on when to use each vs. alternatives. The server reads API keys from environment (good security practice) but provides no guidance on missing keys or degradation paths in tool descriptions.
Doubt every decision in a plan, doubt the doubts, answer each from real evidence — codebase `context` for code doubts, Exa for world doubts — and then RECURSE: doubt each answer in turn (the question behind the question), banking every grounded finding into a knowledge base that feeds the next round and the final plan. Iterate until no new load-bearing doubt remains (convergence). 20 passes is a hard ceiling; recursion is bounded by depth and a global doubt budget.
Exa deep search -> cited, confidence-scored facts (used inside the loop). Low confidence is reported as UNKNOWN and never asserted.
Gate a doubt() result. Proceed only if the plan converged with no open load-bearing doubt remaining; otherwise return the blocking questions.
Input schemas not explicitly visible in source code. doubt() and ground() parameter types inferred from function signatures; verdict() accepts a bare dict with no field documentation.
No output schema documentation. doubt() returns a dict with 10+ fields (passes_used, converged, plan, doubt_log, doubt_tree, knowledge_base, needs_user, engine, note, open_doubts, max_depth_reached) but no schema or field descriptions provided to LLMs.
max_passes parameter lacks constraints. Clamped to 20 in implementation but description does not state min/max bounds or explain the ceiling rationale to LLMs.
No error handling guidance. Tools do not document what happens when API keys are missing, Exa/Fireworks calls fail, or recursion depth is exceeded. LLMs have no recovery path.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
verdict() parameter 'result' is documented as 'the result object from doubt()' but does not specify required/optional fields, structure, or what happens if fields are missing.