Cognitive debt is the next tech debt. Merge Lens turns discarded agent context into explainers, reviewer briefings, and comprehension checks — so you merge faster without approving code nobody understands.
Free forever for public repositories
Files changed
Dependency flow
Comprehension check
Why is the lock acquired before paymentClient.retry() instead of inside the client?
Generated from agent session42 files read2 rejected approachesclaude-code · 11m 04s
The new bottleneck
A 3,000-line agent-authored PR lands. The reviewer can spend hours reconstructing intent from the finished diff — or approve something they won't be able to explain next quarter.
Modern review tools optimize verification. But as agents get better at verifying their own work, the human job changes. Humans need enough context to participate in what gets built next.
The diff is the lossiest possible artifact of what actually happened.
The agent had an entire working session:
At commit time, almost all of it disappears. Traditional review tools only see what survived into the final diff.
Merge Lens captures what the diff throws away.
Agent session
7 artifactsFINAL DIFF
7 of 7 artifacts dropped← traditional AI review tools start here
MERGE LENS
session context retainedHow it works
Four stages, all inside the pull request your team already opens. GitHub stays your source of truth — the Merge Lens dashboard is a focused place to browse repositories, read pull requests, and send AI feedback back to the PR.
Merge Lens records the context generated while Claude Code, Cursor, Codex, or another agent works — before any of it is compressed into a commit.
Instead of walking files alphabetically, the change is organized the way a person actually learns it.
Five free-text questions test whether the author can explain decisions, rather than pattern-match the diff. No multiple choice.
Question 4 / 5 · free text
What breaks if the retry lock is moved inside PaymentClient?
Understanding state is published through GitHub Checks. Nothing replaces the workflow your team already uses.
Required checks are opt-in.
Value by role
Not another AI code reviewer
They solve verification, and they solve it well. We work one layer up: whether the humans on the other side of the merge can explain what shipped.
Architecture
Record what Claude Code, Cursor, Codex, and future agents knew before the final diff existed.
Context competitors cannot reconstruct later
Connect source code to the artifacts that explain it, so questions can be about why the system works this way — not merely what a function does.
“Why does the ledger keep a shadow entry for reversed settlements?”
Grounded in your own history
Turn understanding into infrastructure for reviewer routing, onboarding, ownership, and bus-factor visibility.
person × subsystem × confidence × time
Compounds with every merged PR
Event flow
One GitHub App, one webhook, three generated artifacts, two checks. Nothing to run in CI, nothing to self-host on day one.
Setup path
Design principles
Not every PR needs a comprehension gate. Checks are scoped to changes where losing context actually costs something.
Questions ask for an explanation, not a recognition. You cannot pattern-match your way through them, and we would not learn anything if you could.
Understanding checks do not block merges. They report state, unless your organization explicitly turns them into required checks.
Comprehension state is routing and onboarding signal. It is not an input to reviews, ratings, or headcount decisions.
Comprehension scores must never be used for employee performance evaluation.
If people believe the system evaluates them, they will optimize for the score and destroy the signal. We optimize the workflow around learning, routing, and team resilience instead.
Security
At launch, organizations can run Merge Lens on their own model credentials. Your code travels a data path you already approved, on a bill you already control.
We publish what is decided and mark what is not. No certification is claimed until it has been audited.
Pricing
Comprehension is a per-person property, so billing follows people who actually ship — not diff size, not model spend.
$19–29per active contributor / month
Final launch pricing will fall within this range.
Only contributors with a PR that month are billed.
Open source gets the full explainer and comprehension workflow at no cost. Public code is where lost context does the most long-term damage.
AI code isn't going away
Give your team the context to understand what agents ship — and merge without turning your codebase into a black box.
GitHub-native · BYOK · Free for public repositories