The big picture

Vibestrate runs the AI coding tools you already have. You write the job down once, and a team of AI workers carries it out under rules you set.

Running several models on one job by hand is where the time goes: pasting the same context into a tool that has never seen your project, carrying the plan from one chat to the next, watching each one for drift. Vibestrate is the frame that work happens inside, so every worker starts from the same plan and the same project instructions, written once.

That is the whole idea in one picture: one task, one recipe, one team, one verdict.

A Task is what you ask for, written the way you’d brief a capable colleague. New run at the bottom of the sidebar opens one field for it; Start run begins:

Add structured logging to the settings save handler

Say what you want and leave the how to the workers; naming a file or a rule to respect beats a vague brief. It is the only thing you must provide - everything else has a default.

A Flow is the ordered steps a Task moves through. The Flows page lists this project’s, built-in and your own. The built-in default flow has four steps; Workflow walks through each and the status it produces.

A run doesn’t always take that Flow: with no flow pinned and no project default, Vibestrate sizes up the task and can route a short, low-risk one to a leaner Flow. Flow covers what gets picked, and when.

A Flow never names an AI model. Where a step needs a worker it names the kind, like an implementer, and leaves a labelled empty chair: a Seat. Steps that only run your tests need no chair at all.

A Seat is a spot in the Flow that needs filling, named for the work: an implementer seat, a reviewer seat. The Flow reserves them and says nothing about who sits in them.

That is what makes a Flow shareable: one someone else wrote asks for a planner and a reviewer, naming none of your models or keys. Take one off the hub, drop it in, and it works with your team.

The Crew page settles this before a run does. Flows this crew can run marks each flow Runs, Needs a pick or Cannot start, and the seat map shows which seats are unassigned or have several takers.

Start on an unfilled seat anyway and Vibestrate stops before any work happens, naming the seat and the fix:

This Flow needs the "architect" seat, but crew
"default" has no role that fills it. Open Crew
and add "architect" to a role's Seats.

Two roles filling one seat stops it the same way - the Needs a pick state:

Crew "default" has more than one role filling the
"reviewer" seat (reviewer, senior-reviewer). Pick
one with a role override.

The override is --seat-role reviewer=senior-reviewer on vibe run, repeated per seat.

A Crew is the team that fills those seats. Each member is a Role: a name, a short brief (“you are the Reviewer; you critique the change”), the seats it may sit in, the permissions it runs with, and the Profile it runs at. The Crew page holds all five, one card per role.

Setup writes one Crew called default with six roles:

A Role can fill more than one seat. The Reviewer sits in the reviewer and challenger seats, the Backend Implementer covers implementer, executor and builder. Only the Backend Implementer and the Fixer can write files; the Reviewer can run commands like your test suite but not edit, and the other three run read-only.

Keep more than one Crew and pick which a Task uses, from the compose page’s Crew section or with --crew.

A Profile is how much power you give a worker: which provider, which model, how hard it thinks, and a cap on output tokens. The same Role is cheap or premium depending on the Profile it points at. The Profiles page lists them, and each role card on Crew picks one from its Profile (runtime) dropdown.

This is where you control cost: your best model on the implementer seat, since it writes the real code, something small and fast on a read-only one. Changing a Profile leaves the Flow and the Crew alone.

A Provider is the thing that runs the model. Crew > Providers manages all three kinds:

Set up a provider covers all three.

The defaults work: out of the Setup page you get a Crew with all six roles, one Profile and the built-in Flows, so your first Task needs no seat configured.

One thing is worth changing early: Setup points all six roles at the same Profile, so your reviewer starts out as the same model as your builder. Why you stay in the loop covers what that costs and the one dropdown that fixes it.

Each word has a page of its own: Task, Flow, Seat, Crew.

Your first run → - the same words, watched happening.