Task
❯ ship the new pricing page
Say what you want, the way you'd tell a teammate.
Vibestrate
Give it the whole feature.
Vibestrate is the frame your models work in. It plans the work, breaks out the TODOs, and ticks them off under your rules.
Show me how Vibestrate
Give it the whole feature.
Vibestrate is the frame your models work in. It plans the work, breaks out the TODOs, and ticks them off under your rules.
So no task is allowed to get long.
They all work from one plan. Each item runs in its own session with no baggage from the six before it, and it still knows what was decided and what already landed, because that gets handed forward as real artifacts. You never explain the project twice.
Your rules, not its defaults. The standards you write become policy the run enforces. A supervisor reads every step and decides whether to carry on, send it round again, or stop and come get you.
You make the call at the end. It works to a finished, tested change and waits. If you are not sure where to look, ask Consult. It says what it would check first.
Claude Code·Codex·Gemini·OpenCode·Aider·Cursor CLI·Ollama
On its own, AI loses the plot. With Vibestrate, it ships.
You get a CTO on every step.
It reads the work between steps and gives you its view, good or bad.
Show me howHow a run works.
Six steps, from one line to a finish you approve. Hover a step.
Task
Task
Say what you want, the way you'd tell a teammate. That's the brief.
Flow
Flow
A ready-made recipe of steps (plan, build, review, and so on). Pick one or grab one from the hub.
Crew
Crew
Every step gets a seat. You drop in a model for each, so the right one does the right job.
Run
Run
The steps run in order and each model's work goes straight to the next one. You don't paste anything between tools.
Review
Review
It stops at the risky moments and at the end. You approve, send it back, or let a check decide.
Trace
Trace
Every prompt, change, and dollar spent lands in a log you can read and replay anytime.
And you can ask Consult anything along the way - a read-only advisor that knows your project. It gives advice and can't edit your code.
Read the full walkthroughHow a run works.
One line in, a finish you approve. Swipe through it.
Flow
quality-arbitration · pulled from the hub
Pick a recipe of steps, or grab one from the hub.
Crew
- PlanCClaude
- ReviewGGemini
- ShipXCodex
Drop the right model into every step.
Run
- PlanClaude
- BuildCodex
- TestGemini
Each model's work goes straight to the next one. No pasting between tools.
Review
+128−343 files
It stops at the risky moments. You approve, or send it back.
Trace
- planClaude$0.04
- buildCodex$0.09
- testGemini$0.08
$ vibe run replay a3f2$0.21
Every prompt, change and dollar, logged and replayable.
Same task. Same model.
Four identical runs.
Claude Code, bare
2/4
identical runs shipped stored XSS, and with no gate each one counted as done
0/5
bare builds ever refused anonymous access to everyone’s billing
Claude Code, inside Vibestrate
0
XSS builds ever reached merge-ready - the two that wrote it were stopped at the gate
2/14
builds in the whole benchmark refused the billing read - both were Vibestrate runs
Bare Claude’s best run scored 110 of 120 in two minutes, so the code itself is about as good either way. A Vibestrate run takes 7 to 8 minutes and several model turns instead of one. That buys a plan, a review by a second seat, and a verifier that only signs off on what it could check. A script did the scoring, and all fourteen score files are in the repo.
The whole scoreboard, including the rounds we lost →A hub of flows other people already got working.
Browse flows from people who hit your problem first, pull the one you need, and run it. Publishing yours is one command, under your name.
A flow is a saved recipe of steps. Pull one that already works.
$ vibe flows hub install quality-arbitration
$ vibe flows hub install guy@deep-refactor
$ vibe flows hub publish my-flow
One flow,
any crew.
›Fix the flaky checkout test
Plan Claude Sonnet
Build Codex
Test Gemini Flash
Review Claude Haiku
Fast Crew quick and cheap
Claude SonnetCodexGemini FlashClaude Haiku
One flow. Any crew.
quick · cheap
- Plan Claude Sonnet
- Build Codex
- Test Gemini Flash
- Review Claude Haiku
Free, open, and yours.
Free and open source under Apache 2.0, and light enough to drop into any repo. Read the code, fork it, publish your own flows, and pull the ones other people have already worked out.
One run, start to finish.
Eight steps across five seats. A model hits a rate limit and falls back to a backup mid-run. A reviewer fails and is tolerated. Nothing merges until you have read the diff.
