Ralph Workflow 开源 AI 智能体编排器，用 Ralph 循环跑编码任务
Ralph-Workflow/Ralph-Workflow
Ralph Workflow 是一个免费开源的 AI 智能体编排器，面向编码工作：用户给出一个明确任务，它用选定的编码智能体执行 plan → build → verify → fix 的 Ralph 循环，之后由用户回来检查结果。
Ralph Workflow is a free, open-source AI agent orchestrator for coding work. You give it one well-specified task; it runs a Ralph loop (plan → build → verify → fix) with your chosen coding agent, then you come back to inspect the result. That simple core ships with a strong default workflow for writing software — adopt it as-is first, then extend it when you need to.
Who it's for
Ralph Workflow fits developers and small teams with work that is too big to babysit and too risky to trust blindly. It is not for vague prompts or repos without tests or other guardrails.
Next step
For the shortest honest first run — install, diagnose, one focused task,
and what success looks like — open START_HERE.md.
The fuller tutorial is
Getting started.
The full operator manual lives under
ralph-workflow/docs/sphinx/.
Install from a checkout
For a self-contained manual snapshot, run this from ralph-workflow/ on
Linux, macOS, or WSL with uv 0.7.0 or newer:
make install # `rdev --version` ends in -build # or: make dev # `rdev --version` ends in -dev
Both commands install the rdev launcher only, so they never shadow a
published ralph: an existing global ralph is reported and left in place,
and rdev is used instead of it. An rdev from an earlier dev build is
replaced. Native Windows users should install the published package instead,
for example with pipx install ralph-workflow or pip install ralph-workflow.
make stable owns the global ralph and still asks: continue, remove a pipx
or uv tool installation, or abort. See the
contributor setup guide
for the checkout transaction, failures, stable installs, and switching builds.
Supported agents
Ralph Workflow ships nine built-in agent backends: Claude Code,
Claude Code headless, Codex, OpenCode, Nanocoder, AGY (Google Anti
Gravity), Pi, Cursor, and Kimi (Kimi Code). Pick one, authenticate it once on your
machine, and Ralph Workflow uses it. Selection and trust-boundary
details are in agents and
agent-compatibility.
Documentation route
START_HERE.md— runnable first-run walkthroughdocs/README.md— docs map by reader questionralph-workflow/docs/sphinx/index.rst— maintained operator manual (configure, operate, extend)
Project home
- Runtime: Python ≥ 3.12, local-first.
- License: AGPL-3.0-or-later.
- PyPI: https://pypi.org/project/ralph-workflow/
- Issue tracker: https://github.com/Ralph-Workflow/Ralph-Workflow/issues/new
- Contribution route:
CONTRIBUTING.md→ralph-workflow/CONTRIBUTING.md
The Ralph Loop pattern is attributed to Geoffrey Huntley (ghuntley.com/ralph); Ralph Workflow is an independent reference implementation.
来源：GitHub · github.com