Ralph Workflow is a free, open-source composable loop framework and AI agent orchestrator for substantial, well-specified software engineering work. It gives developers a practical path to vendor-neutral AI coding workflow automation with policy-driven structure and checkpoint/resume confidence: give it a real backlog task before bed, and it plans the work, writes the code, runs your tests, and fixes anything that breaks while your computer stays awake. You wake up to finished code you can open, run, and review.
Plan
Your task description gets tightened before a single line of code runs — gaps and missing expectations are surfaced early so the result actually matches what you asked for.
Build & test
Code gets written, your test suite runs, failures get fixed — all in a single loop with no babysitting from you.
Finished code
Real changed files in your project folder — open them in your editor, run the tests yourself, and decide: merge it, or sharpen the task description and re-run.
Codeberg is the primary repo. GitHub stays available as the mirror.
Best evaluator path: inspect the primary repo on Codeberg, copy a first-task shape, run one real backlog task overnight, then ask one question in the morning: would I merge this?
An example overnight run — before you install anything
From task description to finished, tested code — the example run.
ralphSee the proof on the page — plan, build, and verify — before you install a thing.
Sound familiar?
There’s a better way. Hand off one real task before bed — Ralph Workflow plans it, builds it, tests it, and fixes anything that breaks while the computer stays awake. You come back to code you can review instead of another late-night back-and-forth. No prompts. No babysitting. No midnight sessions. Usually costs less than a dollar in API credits.
The old way
Write task → AI drafts → you check → you prompt → AI tries again → (midnight) → still not done
With Ralph Workflow
Write the task → run Ralph Workflow → keep the computer awake → wake up → open the result → merge or re-run
Simple at the center
Ralph Workflow takes the simple Ralph-loop idea and turns it into a composable loop framework for real software engineering. Every phase is its own loop — plan, build, verify — each one completing cleanly before the next begins. That simple core is what makes the system easier to reason about, extend, and trust. Start with the default workflow and run it as-is. When you need more, extend or reconfigure it — the loop structure stays the same.
This is why Ralph Workflow is an orchestrator and not a coding chat or a thin wrapper. The loop structure keeps planning honest and development disciplined — the result is finished code you can actually open and run.
Planning loop
Tighten the task before any code runs
Gaps and missing expectations get surfaced and closed early — so the result actually matches what you asked for.
Build loop
Write, test, and fix — in the same loop
Code gets written, your test suite runs, and anything that fails gets fixed before the loop hands back — all without you touching the keyboard.
Composable by design
Start with the default. Extend when you need to.
The shipped workflow is already strong for real software work. Build on top of it only when you need to — the simple loop structure stays the same whether you change anything or not.
If this is what you were searching for
Ralph Workflow is a Codeberg-first open-source path for that exact job. If you are looking for Claude Code unattended, Claude Code automation, AI coding workflow automation, or an AI agent workflow composer, this is built for the morning-after question that matters: did you get specific, tested code that is ready to review and honest enough to merge?
Fastest proof path
How you get there
Three phases run automatically, back to back — each one a disciplined loop that completes cleanly before the next begins. No manual steps, no gaps for things to slip through. Plan until it is concrete, build until the tests pass, then stop and hand it back.
Your task description gets pressure-tested before a single line of code runs. Gaps get surfaced and closed. The clearer your description, the better the result — and the easier it is to judge in the morning.
If the request is weak: the work stays in planning until it is specific enough to build and honest enough to judge. A clear done condition is the difference between a useful result and a confusing one.
Code gets written, your test suite runs automatically, and anything that fails gets fixed in the same loop. You don’t see a broken result — the loop handles it before handing back.
If the change is weak: it gets fixed before it moves on.
You get real, focused code changes committed to your project — tested, traceable, and ready to open in your editor. The work is done; the decision is yours.
What you open in the morning: committed, tested code — run the tests yourself and decide whether to merge.
Why it matters
Most AI tools hand you a draft and wait for your next prompt. Ralph Workflow keeps going — plans, builds, tests, and fixes — and hands back finished code you can actually open and judge. Discipline and leverage, not magic.
vs basic AI tools
You get a result that already ran its own tests — not a half-done draft that needs three more rounds of prompting to become something mergeable.
vs hovering over every step
You are completely free to step away — no watching the terminal, no answering prompts. The work runs unattended from start to finish.
vs managing it yourself
The composable-loop structure keeps planning honest and the build disciplined — so the result is not just more AI output, but something grounded in a spec, verified by tests, and traceable in your project.
Already comparing tools you know?
Start tonight. Judge tomorrow.
The best first task is a meaningful one: big enough to matter, specific enough to describe in a paragraph, and familiar enough to judge the result.
1. Pick a real first task
Start smaller than your eventual use case. Integration tests, one feature slice, or one bounded refactor are ideal first runs because the morning-after judgment is cheap and honest. After that, Ralph Workflow can graduate to the bigger feature, migration, or repo-scale cleanup. One clear “done” condition = one clean judgment call the next morning.
2. Write down the job
One clear paragraph is enough. The clearer your description, the cleaner the result — and the easier it is to judge in the morning.
3. Judge the morning-after result
If you would merge it, give it a harder task next time. If not, rewrite your task description and re-run — the second attempt is almost always better because you now know exactly where the gap was.
The one question that matters
Would you merge it? If yes — hand it something that matters. If no — you now know exactly what to sharpen. Either way, one overnight run taught you something real.
pipx install ralph-workflow
Works with the tools you already trust
Plug in the agent you already use — Claude Code, Codex CLI, or OpenCode. Ralph Workflow handles the workflow with a policy-defined loop, so your tools stay the same while the handoff gets more reliable. Resume interrupted runs, swap providers without rewriting your process, and review something you can actually open, run, and merge.
Trust boundary
Keep your existing agent setup. Keep your keys to yourself. Ralph Workflow works through the coding agents you already trust instead of turning “hand over your API keys” into the main setup story.
Why people try it
Runs on your machine
Local-first. Your code never leaves your environment.
Checkpoint & resume
Never lose progress, even if your machine restarts.
Vendor-neutral
Bring your existing agents — Claude Code, Codex CLI, OpenCode.
Anthropic — Claude Code
Strong planning and long context — produces cleaner, review-ready changes.
Codex CLI & GPT
A solid choice for OpenAI-first teams. Works with Codex CLI and the wider GPT model family.
Multi-provider gateway
Routes jobs across multiple providers — swap models without reconfiguring your entire workflow.
No new tools required — use the agent you already trust. The only question is whether the result meets your standard for done.
FAQ
Cost & privacy
Getting started
Results & expectations
Start tonight
Install tonight. Pick one substantial task. Review the result by morning.
Free to install · uses your existing AI agent · your code never leaves your machine
Best evaluator path: inspect the primary repo on Codeberg, pick one bounded task, then install and run it tonight. GitHub stays available only as the mirror.
Free to install. No account needed. Your AI provider’s API is the only cost.