Show HN: Quibbler – A critic for your coding agent that learns what you want (github.com)

Quibbler is a background “critic” for coding agents that automatically watches, reviews, and corrects agent behavior. Announced as an installable tool (pip/uv), it can either hook into Claude Code’s event system or use the Model Context Protocol (MCP) for broader agent compatibility. Instead of repeatedly prompting the agent, Quibbler observes tool use, reads actual changed files, validates claims, checks for tests and verifications, and interrupts or writes feedback when the agent fabricates results, skips tests, deviates from style, hallucinates metrics, or makes changes misaligned with user intent. Over time it learns project-specific rules and enforces them, saving rules to .quibbler/rules.md and maintaining a persistent per-project reviewer. Technically, Quibbler offers two modes: Hook Mode (Claude Code-specific) forwards events to a local Quibbler server and injects feedback via file writes/notify hooks; MCP Mode provides synchronous review through a review_code tool call (requiring user_instructions, agent_plan, and project_path) and a simple MCP server entry. It inspects diffs using a Read tool, returns immediate feedback or approval, and keeps session feedback in .quibbler/{session_id}.txt. Default model is Claude Haiku 4.5 (configurable), and setup requires an ANTHROPIC_API_KEY. For teams using LLM-driven coding agents, Quibbler automates guardrails and institutionalizes review patterns, reducing hallucinations and drift while enforcing tests and style.