This is a metacognitive toolkit for AI coding agents that need to break out of loops without human intervention. It exposes seven core MCP tools like logic-is-too-complex, bug-fix-always-failed, and missing-requirements, plus extended scenarios accessible via list_scenarios and get_prompt. Each tool returns a structured markdown protocol with concrete steps to re-anchor goals, summarize what failed, and pivot strategy. It's built for agentic environments like Cursor, Windsurf, and Claude Desktop where an agent might spiral on circular reasoning or repeated failed attempts. The prompts live in markdown files under prompts/core and prompts/extended, making it straightforward to contribute new recovery patterns. Hosted on Cloudflare Workers, so no local setup required.
claude mcp add --transport stdio io.github.askman-dev-agent-never-give-up uvx agent-never-give-up