openclaw-output-vetter-mcp
JSON →MCP server for verifying AI agent claims vs reality â inline grounding-check, swallowed-exception detection, multi-turn contradiction detection. Catches the agent says 'done' but nothing changed pattern.
Install
pip install openclaw-output-vetter-mcp Tools · 3
- verify_response_grounding Checks whether claims in an AI agent's response are grounded in the input context. Returns verdict (FABRICATED or GROUNDED), ungrounded claim count, overall grounding score, and list of ungrounded claims with overlap scores.
- find_swallowed_exceptions Scans agent-written code for swallowed exception patterns where exceptions are caught and mock/substitute data is returned instead. Returns verdict and findings with severity levels.
- verify_action_outcome Verifies whether an agent's claimed action outcome matches the actual state change by comparing before and after snapshots. Detects state violations of constraints (e.g., read-only mode) and unverified completion claims.
★ 1 GitHub stars