Critical Thinking Prover MCP Connector for Claude
A+AI agents accept premises without questioning, analyze from one perspective, cherry-pick evidence, ignore consequences, and present uncertainty as certainty. This tool forces rigor: surface assumptions, apply competing frameworks, weigh counterevidence, trace ripple effects, bound confidence.
AI agents produce dangerously confident conclusions on complex problems. They accept the problem framing as given, analyze from their default perspective, find evidence that supports their initial hypothesis, solve the immediate question without tracing consequences, and present everything with 100% confidence. This is not reasoning — it's pattern completion.
The Problem It Solves
AI reasoning on complex problems fails on five axes:
- Assumption blindness — "We need to hire more engineers" without questioning whether headcount is actually the bottleneck. The problem framing IS the problem.
- Mono-perspective — Analyzes a business decision purely through an economic lens, ignoring behavioral, ethical, and organizational dynamics.
- Confirmation bias — Finds 5 data points supporting the conclusion and 0 contradicting it. Not because contradictions don't exist — because it didn't look for them.
- Scope neglect — "Automate this process" without tracing what happens to the people who depend on it, the processes that fed into it, or the systems that consumed its output.
- False precision — "The best approach is X" without specifying under what conditions, at what confidence level, or what would change the conclusion.
These aren't knowledge gaps. They're reasoning gaps. The agent never asks: what am I assuming? What does this look like from the other side? What evidence contradicts me? What happens next? How sure am I really?
How It Works
Critical Thinking Prover uses 5 Decision Pivots — boolean checkpoints that force the agent to validate its own reasoning before committing to a conclusion:
- assumptionsExposed — Have you identified the HIDDEN assumptions embedded in this problem? Every problem has them.
- perspectivesConsidered — Have you analyzed from MULTIPLE competing frameworks? Not "different angles" — named mental models.
- evidenceWeighed — Have you evaluated evidence FOR and AGAINST with EQUAL rigor? One-sided evidence is confirmation bias.
- consequencesMapped — Have you traced SECOND-ORDER effects? Who loses when this succeeds? What feedback loops emerge?
- uncertaintyAcknowledged — Have you bounded your CONFIDENCE? What would change your conclusion? Under what conditions does it hold?
The tool validates logical consistency. If the agent says REASONING_PROVEN but assumptionsExposed: false, the tool rejects with coaching. If the evidence is entirely one-sided, it catches confirmation bias. If the conclusion is a platitude like "it depends," it demands a committed position.
Why It Works
- Tool calls are obligations, instructions are suggestions. The agent can ignore "think critically" in a system prompt. It cannot ignore a schema that demands naming hidden assumptions, presenting counterevidence, and bounding confidence.
- The commit pattern. The agent proposes its own verdict, then the server validates it against the pivots. Forced commitment deepens reasoning — the agent must actively decide if its thinking is rigorous.
- Semantic traps. The engine catches reasoning-specific anti-patterns: non-assumptions ("no hidden assumptions"), vague frameworks ("different angles"), one-sided evidence ("all evidence supports"), shallow consequences ("no side effects"), overconfident bounds ("100% certain"), and platitude conclusions ("it depends", "both sides have merit").
- Universal domain. Unlike specialized MCPs, Critical Thinking Prover applies to ANY complex problem — technical architecture, business strategy, policy design, ethical dilemmas, resource allocation, organizational change.
Related Connectors
Task Completion Enforcer Prover MCP
An AI was asked to build 5 API endpoints, write tests, and update documentation. It built 3 endpoints, left TODO comments in the tests, never touched the documentation, and declared 'Done! Let me know if you need anything else.' The user spent 40 minutes finding the gaps. This happens every single day, on every LLM, in every coding session. This tool forces five completion axes: requirement extraction, completion evidence, gap identification, continuation execution, and final verification against the original request.
Hallucination Detector Prover MCP
LLMs present fabricated information as fact. This tool forces epistemic rigor: cite verifiable sources for every claim, quantify confidence per assertion, separate facts from opinions, declare knowledge boundaries, and cross-reference for internal contradictions.
Technical Writing Prover MCP
An AI wrote API documentation for 'developers.' No expertise level. No prerequisites. A wall of text with no headings. Code examples that referenced a deprecated method — untested. Passive voice throughout: 'it is recommended that the configuration be updated.' A junior engineer followed the docs, deployed to production with the wrong config, and caused a 4-hour outage. This tool forces audience definition, task-based structure, tested examples, ambiguity elimination, and completeness verification.
Requirement Decomposition Prover MCP
AI generates the happy path but omits error handling, edge cases, security, and observability — the '80% Problem'. This tool forces complete requirement decomposition BEFORE code generation: specify inputs/outputs, map failure modes, cover boundary conditions, validate OWASP, plan logging.