AI agents wear hats and loop until the job is done

What it does

Ralph is a Rust-based CLI that supervises AI coding agents—Claude Code, Gemini CLI, Codex, and several others—by keeping them in a persistent loop. It plans features, delegates to specialized personas, and re-runs work that fails tests, lint, or type checks. The loop only exits when the agent emits LOOP_COMPLETE or hits a hard iteration cap.

The interesting bit

The “hat system” assigns roles like code-assist, debug, and research that coordinate via events, while a Telegram bot called RObot lets humans block the loop mid-flight to answer questions or steer the agent. It is essentially a control plane for agents that need adult supervision.

Key highlights

• Multi-backend support: Claude Code, Kiro, Gemini CLI, Codex, Amp, Copilot CLI, and OpenCode
• Backpressure gates reject incomplete work based on tests, lint, or typecheck
• Persistent memories and task tracking across loops
• Built-in MCP server mode for integration with MCP-compatible clients
• Web dashboard (Alpha) for monitoring loops via a Rust RPC API and frontend

Caveats

• The web dashboard is explicitly alpha, with rough edges and breaking changes expected
• Running the dashboard requires both a Rust toolchain and Node.js >= 18
• Each MCP server instance is locked to a single workspace root, so multi-repo setups need multiple processes

Verdict

Worth a look if you are already using AI coding CLIs and need guardrails to stop agents from declaring victory too early. Skip it if you want a polished, batteries-included IDE experience rather than a terminal-first orchestration layer.