# Cross-Check AI Outputs for Factual Truth. AI Agent Connect > Agent Hallucination Cross-Checker is a deterministic engine that audits multiple AI agent outputs. It measures consensus, source validity, and semantic contradictions to find out which claims are reliable. Instead of trusting a single agent, you feed the outputs into this MCP. It calculates fact consistency scores, detects contradictions between high-confidence claims, and determines the probability of hallucinations. You can distinguish between total agreement, partial agreement, and outright contradictions, ensuring your AI interactions are high-fidelity. ## Overview - **Category:** audit - **Price:** Free - **Endpoint:** https://edge.vinkius.com/vk_preview_gNakA7A2khcr2iZRVhwDq7FmQPxkJEuT4qaJkwaU/ai-agent-connect - **Tags:** hallucination, consensus, verification, audit, reliability ## Description When you rely on multiple AI agents for complex information, you can't afford to trust just one source. This MCP solves that problem. It’s a deterministic verification engine designed to audit several agent outputs simultaneously. You give it the claims, the confidence scores, and the sources, and it tells you exactly how reliable the information is. It doesn't just tell you if things match; it calculates the degree of agreement. You can pinpoint contradictions between high-confidence claims and get a score for overall fact consistency. This means you move past guessing and get verifiable data, knowing whether your agents are in total agreement, only partially aligned, or outright contradicting each other. ## Tools ### analyze_agreement_depth Breaks down the nature of agent interactions to distinguish between total agreement, partial agreement, and contradictions ### detect_hallucinations Identifies specific claims that are likely to be hallucinations based on probabilistic modeling and contradiction detection ### verify_claim_consensus Calculates the overall reliability of the provided agent outputs through consistency and source metrics ## Prompt Examples **Prompt:** ``` Analyze these agent outputs for consistency: [{'agentId': 'A', 'claim': 'The sky is blue', 'confidence': 0.9, 'source': 'ref_1'}, {'agentId': 'B', 'claim': 'The sky is blue', 'confidence': 0.8, 'source': 'ref_1'}] ``` **Response:** ``` The fact consistency score is 1.0, as both agents provided identical claims with valid sources. ``` **Prompt:** ``` Check for contradictions in these claims: [{'agentId': 'A', 'claim': 'The temperature is rising', 'confidence': 0.9, 'source': 'ref_1'}, {'agentId': 'B', 'claim': 'The temperature is falling', 'confidence': 0.85, 'source': 'ref_1'}] ``` **Response:** ``` A contradiction was detected because the claims are semantic opposites and both agents reported confidence above 0.7. ``` **Prompt:** ``` Calculate the consensus for these outputs with a minimum quorum of 2. ``` **Response:** ``` The consensus report shows that the required quorum has been met for the verified facts. ``` ## Capabilities ### Check Factual Consistency Your agent uses this when it needs to calculate how reliable a set of claims is based on source metrics. ### Identify Contradictions Use this when you suspect two or more agents are making opposing statements about the same topic. ### Measure Agreement Depth This is used to distinguish if multiple agents agree fully, agree partially, or disagree entirely. ### Detect Hallucinations The MCP runs this when it needs to flag specific claims that are statistically unlikely or unsupported. ### Determine Consensus Quorum You use this to confirm if enough agents have provided enough verifiable facts to draw a conclusion. ## Use Cases ### Cross-Checking Legal Summaries You run three different agents on a legal document. This MCP tells you if they contradict each other on key definitions or dates. ### Academic Literature Review Multiple agents summarize research papers. You use the MCP to verify if all agents agree on the primary findings or if there are conflicting theories. ### Market Trend Analysis You feed the MCP outputs from agents tracking different market sectors. It verifies if the consensus on growth rates is consistent across all sources. ### Internal Policy Drafting When multiple departments use AI to draft policy sections, this MCP checks for semantic contradictions to ensure the final document is cohesive. ## Benefits - It calculates fact consistency scores, moving beyond simple agreement counts. - It pinpoints contradictions between high-confidence claims, saving review time. - It determines the probability of hallucinations, giving you a risk assessment. - It distinguishes between total agreement, partial agreement, and outright contradictions. ## How It Works Connecting this MCP is simple. You connect your preferred AI client, and then you feed it multiple agent outputs, along with their confidence scores and sources. The MCP processes the data and returns a detailed audit report. 1. Connect your AI client to the Vinkius catalog and select this MCP. 2. Provide the MCP with the multiple agent outputs you need to audit. 3. Specify what you want to check, such as consensus or contradictions. 4. The MCP runs the audit, calculating scores and flagging inconsistencies. 5. You receive a clear report detailing the agreement depth and hallucination probability. ## Frequently Asked Questions **What is the difference between consensus and agreement depth?** Consensus verifies the overall reliability of the outputs using consistency and source metrics. Agreement depth is more granular; it tells you if the agents are in total agreement, only partially aligned, or if they contradict each other. **Can this MCP detect if an agent is lying?** It doesn't detect intent, but it detects the output. It identifies specific claims that are likely hallucinations based on probabilistic modeling and contradiction detection. **Do I need to provide sources for the agents' claims?** Yes. The MCP uses source validity and confidence metrics. Providing sources allows it to calculate the overall reliability score accurately. **What kind of data does this MCP audit?** It audits textual claims provided by multiple agents. It focuses on measuring semantic contradictions and factual consistency scores. **Is this MCP compatible with all AI clients?** Yes. Because it's hosted on Vinkius, you connect it once from any MCP-compatible client, including Claude, Cursor, Windsurf, and VS Code.