Review Capsule
Join waitlist
Capsules extracted locallyYour code never leaves your machineZero servers, zero infraBuilt for reviewers, not the algorithmWorks with Claude Code & CursorLocal-first by design
Private alpha · onboarding small teams

AI writes code
in minutes.You still review it
like it's 2010.

We attach the author's reasoning to every pull request and re-represent the change as behavior — rendered UI, request/response shape, test assertions. Not diffs. Reach a verdict in sixty seconds, not sixty minutes.

No spam. One email when we open access.
Local-first, zero infraClaude Code & CursorYour code never leaves your machine

✦ The problem

The cost of writing code dropped to zero.
The cost of reviewing it didn't move.

Diffs are a terrible compression format

Reviewers face 40+ changed files with no narrative. Reading a diff to understand behavior is like reading bytecode to understand a feature.

The author's reasoning vanishes

The agent session had every decision, every alternative considered, every uncertainty. By the time the PR opens, all of it is gone.

Risky changes hide in cosmetic noise

An auth boundary moved in line 47 of file 23, surrounded by formatting. Reviewers approve what they can't see. Production finds out.

✦ How it works

Three steps. Zero servers.
Your code never leaves the machine.

Step 1 · Capture

The author's session, attached to the PR

The full agent conversation — what was asked, what was decided, what was rejected, where they were uncertain — travels with the change. Captured locally, scrubbed of secrets, never uploaded to a server.

Step 2 · Surface

Behavior on top, code on demand

Tabs for FE, BE, and tests — each showing the consumer view first: rendered UI scrubbers, request/response shapes, assertion checklists. Diffs only appear if you ask for them.

Step 3 · Score

Hygiene + severity, not just "approved"

Was there a plan? Were options compared? Is rollback documented? Six ownership checks plus a severity model that combines blast radius, AI confidence, and the bits diffs hide.

60s
from PR open to first verdict
0
servers to operate
1
click to open a review
100%
of capture runs locally

✦ See it in action

Same PR. Four lenses.
Sixty seconds to verdict.

A real PR view for “Migrate auth from passwords to magic links” — touches 3 frontend files, 2 backend handlers, and 4 end-to-end specs. Click the persona tabs to switch lenses. The scrubber on the FE tab is draggable.

localhost/repos/acme-app/pr-2026-05-24-magic-link-auth
⌘K
NK
Migrate auth from passwords to magic links
feat/magic-link-auth → main·nikhil·9 files+418−231
1 critical2 hygiene gapsAI confidence: med
Review Capsulev1 · from chat
Intent

Replace password-based authentication with a magic-link flow: users enter their email, receive a one-time signed URL, and exchange it for a session token. Closes SEC-1142.

Decisions made
  • Chose stateless JWT verification over server-side token table — picked simplicity over revocation control. why
  • Token expires in 10 minutes (vs. industry default of 15). why
  • Rejected: keeping password as a fallback for “account recovery” — agent pushed back, would defeat the security goal. why
Open uncertainty
Author was unsure whether /auth/verify-token should invalidate active sessions or coexist with them.
chat turn t-42 · region: handler/auth/verify.rs
Ownership & Hygiene1 critical
Block on the rollback gap — every live session breaks at deploy.
Plan stated before coding
Agent proposed approach in turn t-3 before first edit
Options compared
3 approaches discussed (JWT, opaque tokens, session table)
Diff scope reasonable
4 concerns mixed: auth · UI · tests · config
AI explained assumptions
All 6 questions answered in chat
Rollback steps documented
No plan for active sessions when password column is dropped
Author confidence
One unresolved uncertainty (see capsule)

✦ What you get

Built for the reviewers,
not the algorithm.

The product, in one line

Behavior-first persona views

Frontend reviewers see screenshot scrubbers. Backend reviewers see request/response shapes. QA sees assertion checklists. Code is one click away — never the first thing on screen.

Overview
FE3
BE2
QA4

Review Capsule

Intent · decisions · alternatives rejected · uncertainties — extracted from your agent's session and cited back to the source turn.

Hygiene scorecard

Six ownership checks — was there a plan, were options compared, is rollback documented — surfaced at review time, not write time.

Severity you can audit

Every finding shows blast radius, AI confidence, coverage delta, domain sensitivity. No black-box "looks risky."

Keyboard-first, < 1s paint

1·2·3 switch persona · j·k walk findings · / search · ? help. Built for reviewers who live in their editor.

Local-first, zero infra

Capsules, screenshots, severity — all computed on the dev's machine. Your source never leaves. No cloud account.

✦ FAQ

Answers to the obvious questions.

Stop reviewing diffs.
Start reviewing behavior.

Get an invite when we open access. We're onboarding small teams first — tell us about yours when you sign up.

No spam. One email when we open access.
No credit card requiredOne email when invites openUnsubscribe in one click