Early access is invite-onlyJoin the list →

What should your agents finish tonight?

Hand off a coding task before bed. Agents work through the night, your tests decide when it's done, and a receipt is waiting at breakfast.

Good morning · your-app
Add CSV export to the orders pageYour model
  1. Builder

    Done, and it's fully tested.

    sent back
  2. Jarvai

    Hold on. Tests claimed, no runner called.Sent back to Builder.

  3. Builder

    npm test

    orders › exports a header row

    orders › escapes commas and quotes

    orders › streams 10,000 rows

    orders › keeps ISO dates

    12 passed · 0 failed

  4. Jarvai

    Checks passed. Protected tests untouched.

Good morning. 3 tasks ready for review.

Ideas for tonight

Hand off anything your tests can judge.

Pick one and watch it run.

How it works

One night with Jarvai

  1. Hand-offHand off the backlog.

    Queue tasks from GitHub, Linear, Slack or your terminal. Each gets its own agent, its own sandbox and a finish line: a command that must exit 0.

  2. Night ShiftAgents work in parallel.

    Six at once, each on a fresh copy of your repo, each with a hard budget it can't overspend.

  3. FlagOne says done. It isn't.

    Done means your command exits 0 on a clean checkout. A claim without that goes back to the agent.

  4. RaceThree models, one task.

    Race Claude, GPT and your own model on the same task. The tests pick the winner.

  5. Finish LineCI broke. Fixed before you woke.

    The agent watches its pull request, reads the failing check and fixes the cause.

  6. Morning BriefGood morning.

    Pull requests with signed receipts. Two need a decision from you, and one stopped at its budget and says why.

The vocabulary

One night, eight words.

Everything Jarvai does has a name, and every name means one exact thing.

  1. Hand-off

    A task you give Jarvai before bed, with the command that decides when it's done.

  2. Night Shift

    Every hand-off running at once, each with its own agent and its own sandbox.

  3. Finish Line

    Your command, exit 0, on a clean checkout. Done means this and nothing else.

  4. Flag

    Raised when a claim has no evidence. Flagged work goes back to the agent.

  5. Race

    One task, several models. The tests pick the winner.

  6. Cap

    A hard limit on money, time and turns. Hit it and the run stops.

  7. Receipt

    Signed proof of a run: commit, commands, exit codes, checks, spend and model.

  8. Morning Brief

    Your phone at breakfast: what's verified, and what needs you.

Morning Brief · 07:00

Wake up to proof.

Every pull request arrives with its receipt: the tests that ran, what changed, and a recording of the app doing the job. Review it over coffee and merge from your phone.

  • ✓
    A receipt on every pull requestTests, commands and diffs, printed from what actually ran.
  • ▶
    See it workingA short recording of the change running in a real browser.
  • ?
    Only the real questionsAgents ask when a call is yours to make, and wait for your answer.

Trust

Agents don't get to grade their own work.

Finish Line

Tests the agent can't touch.

Done means your command exits 0 on a clean checkout. Your tests are locked: if an agent edits one, Jarvai puts the original back before the check runs.

tests/export.spec.tsprotected
1test("exports every order", async () => {2  const rows = await exportCsv(orders)3  expect(rows).toHaveLength(12)4})

The check ran against your original test.

Edit undone and flagged

Flags

It catches a done that isn't.

Every claim is checked against what actually ran. No evidence, no done.

Builder

All done. Every test passes.

Jarvai

Hold on. No test output was recorded this turn. Sent back.

Builder

Ran npm test: 12 passed

Jarvai

Verified. Now it's done.

Receipts

A receipt for every turn.

Printed from what happened. When nothing was checked, it says so.

Receiptturn 9
edited
2 files · +52 −4
ran
5 · all exit 0
checks
12 passed
spend
$0.62 of $2
commit
a1f3c9e · signed
checks passed

Every approval on the record.

Choose what runs on its own and what waits for you. Approvals are chained, so a tampered log is rejected.

Undo any write.

Every file is saved before an agent changes it. Drag back to any of the last fifty turns.

turn 1undo to turn 37turn 50

Caps

A budget it can't overspend.

Every task gets a hard cap on money, time and turns. Hit it and the run stops, with the receipt saying why.

Speed up the search query$2.00
cap $2.00

Stopped at the cap after 41 turns. Receipt explains what was tried.

Any model

Bring any agent. Keep your rules.

Jarvai speaks the OpenAI-compatible API format. Plug in Claude, GPT, Gemini or a model you run yourself, switch any night, or race them on the same task. Run a local model and your code never leaves your machine.

Why we built Jarvai

Our first agent team approved its own work.

One sentence became a planner, a lead, parallel workers and a reviewer that decided when to ship. The reviewer approved 23 of 53 runs.

Then we checked them the strict way: a person accepted the result, nobody fixed it by hand, and a check outside the agents confirmed it. Not one run qualified. Jarvai exists to close that gap.

Approved by its own reviewer23 /53
Actually shipped0 /53

Our run log, 19 April 2026

  1. First agent team
  2. A strict definition of shipped
  3. Two builders, one workspace
  4. Receipts
  5. Tamper-evident approvals
  6. Early access

FAQ

Questions

How do I get access?

Early access is invite-only. Join the list below and tell us what you'd hand off first. We'll email you when your invite is ready.

Which agents can I use?

Any agent you like. Jarvai speaks the OpenAI-compatible API format, so Claude, GPT, Gemini and your own models all plug in. The rules stay the same for every agent.

How many tasks can run in one night?

As many as you queue. Every task gets its own agent and its own sandbox, so a night can cover a whole backlog. You decide how many run at once and how much they may spend.

Where does the work happen?

On your machine, in an isolated copy of your repository, so your branch stays untouched until you approve the changes. The code an agent reads goes only to the model provider you choose. Run a local model and nothing leaves your machine.

What is on a receipt?

The files that changed, the commands that ran, the checks that passed or failed, and anything the agent was refused. When no check ran, the receipt says so in plain words.

What if I don't have tests?

Then the receipt will say that nothing was checked, and the result stays the model's claim. Writing the missing tests is a good first task to hand off.

Hand off tonight. Read the receipt at breakfast.

Join the early access list. We're inviting a small group first.

We'll only use your email to send your invite. Prefer email? founder@jarvai.dev