Solves the messy LLM output problem when you need reliable JSON back from a model. Exposes three tools: extract_json pulls JSON from prose or fenced blocks using fallback strategies, validate_response gates parsed values against a shape spec with required and optional fields, and build_retry_prompt generates the exact feedback message to send back when validation fails. Built on the agentcast library. Drop it into Claude Desktop, Cursor, or any MCP client via npx. Useful when you're building agents that parse structured responses and need robust extraction plus a standardized retry loop instead of brittle regex or silent failures.
claude mcp add --transport stdio io.github.mukundakatta-agentcast uvx agentcast