Product
Every agent is a cursor you can watch.
quibbit opens real browsers on your machine and seats a named crew in them. Each agent carries a pill that says what it is doing. Steer one, take over, or answer when it asks you for a code.
Free plan for one person. Windows build on the Download page.
Plan a run
Pick the flows. Seat the crew. Choose a size.
A run starts with a project and the flows you want covered. Seat the agents whose job fits: Pip takes checkout, Otto takes sign-in, Mabel takes the forms.Dex takes checkout, Moss takes sign-in, Nell takes the forms. Pick Quick, Standard, or Deep, read the estimate, and start.
Test accounts come from the vault. Each agent gets its own clean browser session on your machine, so one agent's cart never shows up in another's test.
The referee
Three fresh browsers before you hear about it.
An agent's finding is a claim. The ref opens three clean browsers, one at a time, and walks the same steps. If all three reproduce it, the finding lands in your inbox with its video, steps, and network log. If they don't, you never see it.
The ref is best-effort. It catches flaky claims, not every bug, and it tells you the count it saw.
Sections
Twelve sections, each with a specialist.
Turn a section on and its agent joins the run. Every section runs on your machine.
- FlowsPip, Otto, WrenDex, Moss, Rook
- FormsMabelNell
- Console and networkDex
- Design reviewNell
- AccessibilityAda
- PerformanceTally
- ResponsiveBeaRemy
- Cross-browserRemy
- ContentClemZuzu
- RegressionMoss
- ExploratoryJuno, RookTally, Rook
- Pen testFaceless, on verified domains, after approvalComing
Design review
Checks the build against the design
Give the crew a Figma file or your design system. It compares spacing, type, color, and components in the running app, and reports where the build drifted from the design.
Export
Tests you keep
Every flow the crew learns saves as a plain Playwright test. Export them any time, on every plan, and run them in your own CI without quibbit.
Schedule your saved tests nightly or on every deploy. They replay on our browsers while your laptop is closed, with no model calls.
ComingRun sizes and budgets
Runs that know when to stop.
Every run has a size, an estimate before it starts, and a budget it stops at. One run can't drain a subscription or an API budget.
Quick
4 agents × 8 steps
about 0.3M tokens
The first run on a subscription is always Quick, so your first impression isn't a limit warning.
Standard
8 agents × 12 steps
about 1M tokens
A release check on a few flows. The default once you've seen a Quick run finish.
Deep
16 agents × 20 steps
about 4M tokens
Every section on, before a big release.
What a budget does during a run
Before it starts you see a range: dollars on an API key, tokens on a gateway.
Standard · estimate about 1M tokens
Steps are capped per agent. An agent stops when its task is done.
You get a banner in the app. Keep going, or stop and read the report so far.
When agents hit the top of the range, the ref still checks what they found. You get a partial report, not nothing.
If a limit hits mid-run, the run saves its plan, progress, and findings, and closes its browsers so hosted minutes stop. When the limit resets, resume from the last finished task. Nothing replays a step that changed data.
On a Claude subscription the estimate is a guide, not a ceiling, because your Claude Code makes the model calls. quibbit caps its own side: steps, agents, and how much each tool call returns.
See the crew on your own site.
Download, add a model key, and start a Quick run. The first finding comes with its video.
Free for one person. Windows for now.