Claude Code handles one task at a time. PromptWheel orchestrates autonomous agents that scout, execute in parallel, learn across runs, and batch everything into milestone PRs — all as draft PRs you review before merging. It turns a coding assistant into an agent orchestration layer.

Yes — run /plugin marketplace add promptwheel-ai/promptwheel then /plugin install promptwheel-ai@promptwheel. Restart Claude Code, then run /promptwheel:run. It uses your existing subscription, no API key needed.

Instead of creating a separate PR for every fix, milestone mode (--batch-size) accumulates changes on a single branch and opens one PR per batch. An 8-hour run might produce 5 milestone PRs containing 50+ improvements total, instead of 50 individual PRs.

Every change runs through typecheck and tests before merging. Failed tickets are automatically blocked. All PRs are ready-for-review by default, and the trust ladder limits changes to safe categories like refactoring, tests, and docs until you opt in to more.

PromptWheel is free and open source. Four ways to run: (1) Plugin inside Claude Code — uses your subscription, no API key needed. (2) CLI with Claude — requires ANTHROPIC_API_KEY. (3) CLI with Codex — uses codex login or OPENAI_API_KEY (no Anthropic key needed). (4) CLI with OpenAI — requires OPENAI_API_KEY. API costs depend on your codebase size, run duration, and provider pricing.

Formulas are repeatable recipes for specific goals. Run --formula security-audit to focus on vulnerabilities, --formula test-coverage for tests, or --deep for principal-engineer-level architectural review. You can also write your own.

PromptWheel's plugin is modeled after Anthropic's own Ralph Loop plugin — both use the same Stop hook mechanism to maintain a continuous session inside Claude Code. The only addition is an MCP server (following the same npx pattern as the official Firebase and Playwright plugins) for state management. All code execution happens through the official Claude Code CLI with your own credentials — PromptWheel never stores, proxies, or shares API keys. The plugin path uses your Claude Code subscription directly; the CLI path requires your own API key.

Codex CLI provides a built-in exec subcommand designed for non-interactive/automated use, along with safety modes (Read Only, Auto, Full Access) and OS-level sandboxing. PromptWheel uses the official Codex CLI with your own credentials (codex login or OPENAI_API_KEY) — no credential sharing, proxying, or storage. PromptWheel doesn't circumvent rate limits or safety features, and uses --sandbox workspace-write by default for restricted writes. It's the same relationship as a CI pipeline invoking the Codex CLI in a build step.

PromptWheel follows current Codex CLI model restrictions. With codex login (no OPENAI_API_KEY), you can use gpt-5.3-codex (default), gpt-5.2-codex, and gpt-5.1-codex-max. Models like gpt-5.2-codex-high, gpt-5.2-codex-xhigh, gpt-5.1-codex-mini, and gpt-5.2 / gpt-5.2-high / gpt-5.2-xhigh require OPENAI_API_KEY. If your saved model requires API-key auth, PromptWheel prompts you to choose a compatible one.

PromptWheel automatically loads your CLAUDE.md (for Claude runs) or AGENTS.md (for Codex runs) and injects it into every scout and execution prompt. If neither file exists, it generates a baseline from your package.json. During long runs it re-reads the file periodically, so you can update conventions mid-run. You can also point it at a custom file or disable it entirely via config.

PromptWheel works with any codebase that Claude Code supports—JavaScript, TypeScript, Python, Go, Rust, and more. It creates PRs on GitHub, so you need a GitHub remote.