Pre-release · waitlist open
Your agents ship more code than you can review.
Truespec.ai checks every pull request against the ticket it closes, then tells you which ones need you. On the day below, that is 5 of 23.
5 things in here need a person. Which 5?
Needs you · open a row for the evidence
✕ Not readyInvite link 404s for existing users#431 Add guest invitesENG-519 · AC 3
Claude Code · +2,412
- Invite ada@example.com, who already has an account
- Open the invite link while signed out
- Expected the sign-in page. Got a 404 at
/invite/accept
✕ Not readyExport stops at 10,000 rows#428 Export to CSVENG-498 · AC 2
Codex · +1,106
ENG-498 says the export includes every row. It stops at 10,000. This pull request also changedexport.spec.tsto expect 10,000.✕ Not readyAnnual plans show $0.00#436 Billing page copyENG-530 · AC 1
Cursor · +184
On/billing, staging shows $0.00 for annual plans. ENG-530 expects $190.00. Screenshot attached to the pull request.■ Your decisionRewrote the auth middleware, which the ticket never asked for#433 Password reset for guestsENG-512 · 31/31 passed
Claude Code · +904
Every check passes. The pull request also rewritesauth/middleware.ts(+312 −280). ENG-512 only asks for a reset email. Decide whether the rewrite ships with it.■ Your decisionThe ticket doesn't say what happens past 200 lines#439 Invoice PDFENG-533 · spec gap
Codex · +677
The agent cut invoices off at 200 lines. Suggested acceptance criterion: the PDF paginates and repeats the header on each page.
18 ready to ship · merge without reading the diff
- #429 Empty states12/12 passed
- #430 Audit log44/44 passed
- #432 Webhook retries18/18 passed
- #434 Settings nav6/6 passed
- #435 Rate limits22/22 passed
- #437 Search index37/37 passed
- #438 Dates in every locale15/15 passed
- #440 Avatar upload11/11 passed
- #441 Session expiry9/9 passed
- #442 Onboarding checklist14/14 passed
- #443 Team switcher8/8 passed
- #444 CSV import errors19/19 passed
- #445 Email footer4/4 passed
- #446 Usage chart16/16 passed
- #447 API key rotation21/21 passed
- #448 Invite reminder7/7 passed
- #449 Dark mode tables10/10 passed
- #450 Archive projects13/13 passed
The problem
Writing the code stopped being the slow part.
An agent opens a pull request in minutes. Reading it properly still takes your full attention, and by lunch there are 23 of them. Teams end up choosing between three bad options.
A review is supposed to answer one question. Is this what we asked for, and does it work? Reading diffs is a slow way to answer it.
How Truespec.ai works
It reads every line. You read the verdict.
1
Starts from what you meant
Reads the ticket, its acceptance criteria and the decisions behind it, from your tracker, docs, chat and meetings.
2
Checks what was built
Runs the pull request and tests each acceptance criterion. Lists the changes the ticket never asked for.
3
Gives a verdict with evidence
Ready or not, posted on the pull request. A failed check comes with steps to reproduce. You decide what ships.
Use cases
Six moments where agents outrun you.
Pick one. They run in the order they happen, from the ticket to the question someone asks months later.
Gaps found before work starts
An agent builds what the ticket says and fills every gap with a guess. Truespec.ai reads the ticket first, lists what it leaves open and suggests acceptance criteria. Every later check measures against those criteria.
Description
Add a PDF export to invoices.
Acceptance criteria
None.
Truespec.ai · 3 gaps an agent would have to guess
- Page size and what happens past one page
- Currency and date format for non-US customers
- Whether voided invoices can be exported
Suggested: 4 acceptance criteria, drafted from the 9 October billing review.
Add to ticketEdit first
A check before the pull request opens
Coding agents run Truespec.ai on your machine through truespec mcp. The agent looks up the spec, runs a local check and fixes what fails. The pull request it opens has already passed once.
The diff, mapped to the ticket
Each acceptance criterion gets a result and the evidence behind it. Changes the ticket never asked for are listed on their own, so you know which 2 files of 38 to read yourself.
Not ready to ship. 1 of 5 acceptance criteria failed.
| AC 1 | A member can invite a guest by emailinvite sent from Settings › People on staging | ✓ Passed |
| AC 2 | Guests can read and comment, not editedit controls absent; API returns 403 | ✓ Passed |
| AC 3 | Existing users land on the sign-in page404 at /invite/accept · steps to reproduce | ✕ Failed |
| AC 4 | Guests never see other guestschecked lists, mentions and search | ✓ Passed |
| AC 5 | An invite expires after 7 daysinvite_service.ts:88 | ✓ Passed |
Outside the ticket · 2 changes to read yourself
auth/middleware.tsrewritten, +312 −280stripeupgraded from 14.2 to 16.0
A verdict that doesn't rest on the agent's word
The agent's summary is a claim. Truespec.ai runs the pull request on staging and tests each criterion itself. It also tells you when the pull request edited a test to make it pass.
The agent's summary
"Done. Implemented CSV export with streaming. All 214 tests pass."
Truespec.ai's check
Not ready to ship. 1 of 31 checks failed.
✕ Failed The export stops at 10,000 rows. ENG-498 says every row.
export.spec.ts was changed in this pull request to expect 10,000.
One message, sorted by what needs you
Ready pull requests are listed so you can merge them. Each failed one names the criterion it failed and links to steps to reproduce.
Overnight, 3 agents opened 12 pull requests. Truespec.ai checked all 12. 3 need you.
- ✕ Not ready · 2#455 Search returns archived projects ENG-541 AC 2
#462 Reminder email is sent twice ENG-547 AC 1 - ■ Your decision · 1#458 adds a dependency,
pdfkit. ENG-533 doesn't mention one. - ✓ Ready · 9#453 #454 #456 #457 #459 #460 #461 #463 #464
An answer with its sources
Truespec.ai keeps a linked record of your code, tickets, meetings and chat. Ask it why, and the answer names the pull request, the ticket and the decision. A change an agent made in November still has a reason in March.
Why does invite_service.ts retry the mailer three times?
The staging mailer was returning 403s. Claude Code added the retry in PR #431 for ENG-519, on 12 November. Maya Kim approved it in the review thread and asked for the limit of three.
PR #431 · review thread
ENG-519
invite_service.ts:112
Don't want these issues on your pull requests?
Join the waitlist and Truespec.ai will email you when early access opens.
The fair question
Why trust an agent to check agents?
Don't.
Trust the evidence it shows you.
Join the waitlist- 01
It checks against your spec
A verdict maps to acceptance criteria you wrote or approved. Truespec.ai has no opinion about code style to push.
ENG-519 · AC 3 · "Existing users land on the sign-in page"
- 02
Every result opens
Behind each pass or fail is the criterion, the run and the steps to reproduce. You can check any of them in a minute.
✕ Failed404 at /invite/accept · 3 steps to reproduce
- 03
It says how sure it is
When the cause is uncertain it says "Likely cause". When the ticket is silent it says so and asks, instead of guessing.
Likely cause: mailer permission on staging. Not confirmed.
- 04
You decide what ships
Truespec.ai gives the verdict. Merging and releasing stay with you.
Verdict: not ready to ship · Merge: yours
Any author
Truespec.ai checks the pull request, whoever wrote it.
Claude Code
Cursor
Codex
Copilot
- A person
Your tools
Connections bring in the code, tickets, docs, chat and meetings.
GitHub
Linear
Jira
ClickUp
Slack
Teams
Notion
- Meetings
Your region
Each workspace chooses a region when it's created. Its data is stored there.
Europe
United States
Keep the speed. Know what you're shipping.
Join the waitlist. Truespec.ai will email you when early access opens, and with product updates on the way there.