What you can hand it.
The same loop fits very different work: say the outcome, approve the exact terms, let agents do it, and come back to results that prove themselves.
Clear the backlog overnight.
Queue the tasks you never get to before you log off. The crew builds them in parallel, each in its own checkout and branch, and recovers by itself if the machine restarts. In the morning, results wait as Ready with the diff, the check and the screenshots.
Nothing is marked done until you look.
$ toolroll status
Running: 1 — search-index (#431, running checks)
Ready for review: 5 — csv-export (#424), rate-limit (#425),
orders-pagination (#427), flaky-checkout (#428), dark-mode (#430)
Queued: 0
Plan windows: claude — 5-hour 3%, Weekly 21%Decide from your phone.
Scopes, results and drafts reach you in Telegram, Slack, Discord or Teams with the buttons you need: Approve, Edit, Send back. It's the same task as the console, so a tap on the train is the same decision as a click at your desk.
You're only woken for decisions that need a person.
Ready: Rate-limit the public API
check passed · 6 files · codex
Draft to approve: reply to a refund question
Support desk · Maya
Ship behind a real check.
Your project's own command decides what passes. It runs on the exact commit being released, and a deploy refuses anything that wasn't checked and marked complete. Toolroll ships itself this way.
Failed checks stay failed, whoever marks the result.
== summary
unit: exit 0
Test Files 202 passed (202)
flows: exit 0
18 passed, 0 failed · 5.4 min
app: exit 0
lead — 7 passed, 0 failed · 3 min
candidate 6b0e44b · tree matches · deployableTriage issues as they arrive.
Draw your process as zones and let cards flow through it. A new GitHub or Linear issue starts a card, Jev sorts it, real bugs build as ordinary tasks under the same approvals and checks, a person decides what ships, and the issue gets its comment. The engine spends no model tokens moving cards.
New issuesGitHub
SortJev
Buildclaude
Person decidesyou
Update the issueGitHub
AI teammates for the rest of the work.
Starter kits set up a working teammate in one click: its rules, its memory, the flow it works and the buttons that start it. You approve what leaves the building; routine calls inside the limits you set go ahead without you.
Each kit's page lists what's left to connect.
Starter kits
Support deskMaya
Drafts a reply to each customer email. You approve it before it goes out; anything it can't answer comes to you.
Bug triageTheo
Sorts each new issue into bug, question or noise, answers questions, and turns real bugs into fix tasks you approve.
Sales follow-upLeo
Writes back to each new lead. You approve the email; it follows up when there's no answer in three days.
Ops requestsAda
Approves routine requests under $200 without you and brings you everything else.
One lead for a small team.
Teammates join a shared conversation with the same lead, sign in with your identity provider, and see the projects their group allows. Budgets cap spend by project, person or teammate, and a sealed ledger records who approved what, and when.
Every act has a name and a time on it.
Action ledger
- alex approved the scope of
invoices-csv - claude built
invoices-csv · check passed - sam asked for a revision of
invoices-csv - sam marked Complete
invoices-csv
Start with the one you need.
Every use case here is the same install. Look around in the sandbox first, or start it for real.