KeepState checks agent work against acceptance tests you approve before the run. Every attempt and cost stays on the record; each verdict carries a signed receipt.
$10.00 of session and storage credit at signup. Bring your own model key. Model usage is billed separately.Verification and signed receipts today. Receipt-triggered payment is planned.
KeepStateIllustrative receipt
Work, accounted for.
One record. A basis both sides can inspect.
Work
Bounded coding task
Acceptance
Buyer-approved tests
Verifier
Independent host runner
Record
Attempts · cost · verdict
Illustrative verdict✓Accepted
✓Acceptance tests approvedBuyer approves before the run.
✓Checks run independentlyThe agent cannot reach the checking runner.
✓Every attempt accounted forAttempts, cost and verdict recorded.
Basis for acceptanceYour tests decide
ILLUSTRATION ONLY No real job, signature or payment.Buyer-approved tests. Independent verification.
The work has an account. Now give it a receipt.
Cruise · live
Verification and signed receipts. Buyer-approved tests. Host-run checks. Every attempt accounted for.
Clearing · planned
The receipt releases the payment. Receipt-triggered payment is planned. Not live today.
01 / From a job to a shared record
Agree on done. Then let the work begin.
KeepState’s Cruise workflow turns the buyer’s acceptance tests into an independent record of what passed. The agent does the work. The tests decide the outcome.
01
Lock
Start from the repository’s existing checks. Approve the acceptance manifest before the run.
02
Run
The agent works in a durable session. Save points keep its files and context ready to resume.
03
Verify
Pinned tests run on the host, beyond the agent’s reach. Forged pass logs are refused.
04
Escalate
Restore the checkpoint before trying the next model. After the last model, review the job: resume with a wider ladder or cancel.
05
Receipt
Reconcile every attempt and cost. Each verdict carries the locked tests’ digest and a signed receipt.
Cruise supports bounded coding tasks with deterministic tests. Payment clearing is planned.
Cruise checks work with an independent runner the agent cannot reach. It starts with a cheaper model and restores the checkpoint before escalating a failed attempt.
Cruise. Set the task. It drives, tests, and retries.
Current scope: bounded coding tasks (scoped tasks with a fixed test suite) on the KeepState runtime, across Anthropic-native and OpenAI-compatible model families.
Recovery covers supported agents and saved checkpoint state. External effects stay outside that boundary. Current availability →
Verify. Your checks decide what passes.
Recover. Return to recorded state.
Account. Every attempt, on the record.
02 / Continuity beneath the receipt
Your agents never start over.
A failed attempt should not erase the work that came before it. Retries fork from the parked checkpoint. Saved session state supports the next attempt.
One saved stateA considered next step
Workflow illustration. A saved state supports the next verified attempt.
02 / Recover
Your agents never start over.
Rewind. Resume. Keep working.
Checkpoint the running session. Rewind to an earlier save point and resume with its memory and files intact. Work after that save point is lost.
Provider-side model cache is not included in the checkpoint.
A session, resumedRecorded kill-and-resume
keepstate / sessionresumed
$ ks demo --task fixture --kill-at 3m
session a1618683e4bb · coding task
KILL · the agent's machine, hard
back from save point 1 · 25.2 s later
same session · memory and files intact
paid model calls replayed, not rebought
TESTSpython3 test_inventory.py → OK
Same session. Task continued.step 3 of 3
Fork schematic Choose a branch
One saved point
Kill-and-resume, recorded 29 Jul 2026 (the gate-4 demo): three minutes into a coding task, the agent's machine and its supervisor were killed on purpose; the same session was back at work 25.2 s later and its test passed. Editorial summary. View the unchanged original record.
Go back to a saved state.Read the recorded kill-and-resume
Recorded test·the kill and the resume, from gate 4 · edited for readability; the record is verbatim·original record ↗
Recorded on 2026-07-29 on the Apple Silicon evidence bench.
A coding session was checkpointed, interrupted by a hard kill, and resumed.
The recorded recovery completed with the workspace and session state preserved.
10 model calls were recorded; 29,871 replayed tokens were not purchased again.
The recorded workload check passed. These are historical test results, not a recovery-time guarantee.
Recorded test·rewind, from the plaintext-store fleet test·original record ↗
See which capabilities are proven, which are enabled in production, and what KeepState does not claim.
Cruise requires bounded coding tasks with deterministic tests. Customer data requires classification and named authorization; the manifest enforces what you supply and does not detect sensitivity.
The vendor’s account of its own work is only one side of the story. KeepState gives the buyer and vendor a record grounded in tests agreed before the work begins.
Tests are locked on approval. The verifier refuses changes to the pinned test code.
02
The verifier sits outside.
Host-run checks are out of the agent’s reach. A fabricated pass log is not a pass.
03
The receipt keeps the account.
Attempts and costs stay in the ledger. A signed verdict receipt makes later changes detectable.
Accounting record and pricing
03 / Account
What did every attempt cost?
KeepState records attempts and costs in its ledger. In the accepted-work test, settlement cleared only after every attempt reconciled, including failed attempts.
Publication authorization only: this scope does not create an external pull request.
The recorded fixture uses a placeholder tariff. It is not a customer price.
KeepState runs your AI agents under your own acceptance rules, on save points it can rewind and retry, with a ledger of reconciled costs.
Run coding agents with saved checkpoints, independent checks and an account of every attempt.
Run AI agents with rules for done.
KeepState runs your agents, checks their work against your acceptance rules, and keeps save points for recovery. Its ledger accounts for attempts and costs.
Available to every hosted account. Cruise supports bounded coding tasks with deterministic tests.
Sign up with no card for $10.00 in credit, once per identity. Credit covers session time and checkpoint storage; model tokens use your own provider key.