Swap Claude, Cursor, and Codex like batteries until the PRD is done

What it does

Ralphy is an orchestration wrapper around existing AI coding CLIs—Claude Code, Cursor, Codex, OpenCode, Qwen, and others. You point it at a task list or PRD, pick an engine, and it runs the agent in a loop, retrying and switching contexts until the tasks are marked complete. It handles the bookkeeping: tracking progress, isolating parallel runs in worktrees or lightweight sandboxes, and merging results back into your branch.

The interesting bit

The project treats AI agents as interchangeable labor rather than oracles. It abstracts away the per-engine permission flags and output formats, and its sandbox mode symlinks heavy dependency folders instead of copying them, which keeps parallel runs from drowning in node_modules. That orchestration layer—managing branches, stashing local changes, and resolving merge conflicts—is the actual product.

Key highlights

• Supports eight AI engines (Claude, OpenCode, Cursor, Codex, Qwen, Droid, Copilot, Gemini) with model overrides and engine-specific argument passthrough.
• Reads tasks from Markdown PRDs, YAML, JSON, or live GitHub Issues, including parallel group ordering.
• Parallel mode spins up isolated agents with their own branches; can auto-merge, open PRs, or leave branches untouched.
• Sandbox mode uses symlinks for read-only dependencies to avoid duplicating gigabytes across parallel worktrees.
• Optional browser automation via agent-browser for UI testing, plus webhook notifications for Discord or Slack.

Caveats

• Requires the underlying AI CLI tools to already be installed and authenticated; Ralphy is strictly glue and will fail if the engine CLI errors out.
• The bash version needs external tools like jq and optionally yq, while the npm version needs Node 18+ or Bun.
• Parallel execution reliability depends on git worktree behavior, and the README notes automatic fallback to sandbox mode when worktrees fail (e.g., nested repos).

Verdict

Worth a look if you are already using AI coding agents and want a foreman to run them in shifts against a spec. Skip it if you do not have the underlying CLIs installed or you prefer to pair-program manually with a single model.