Now onboarding a small group of design partners. Request access →
MergeProof Request access

VERIFICATION FOR AGENT-WRITTEN CODE

Merge agent pull requests on evidence, not on a skim of the diff.

MergeProof agrees what “done” means before the agent starts, then independently verifies the running application against it and attaches signed evidence to the pull request.

Request access How verification works
app.mergeproof.ai / storefront / pull / 482

PR #482 · ENG-214 · agent: claude-code

Add discount codes to checkout

Blocked · 1 criterion failed
Proven

A valid code (SAVE10) takes 10% off the subtotal

Browser · $84.00 → $75.60 · screenshot, trace

Proven

An invalid code shows an error; total unchanged

Browser + API · 422 returned · request log

Failed

An expired code is rejected with “This code has expired”

Expected total $84.00 · observed $75.60 · added from bug BUG-31

Proven

The discount is applied before tax

API · tax computed on $75.60

Proven

The code field is labelled and keyboard-usable

Accessibility scan · 0 violations

Sign-off

Error and success copy fits the brand’s tone

Human review · awaiting product owner

EVIDENCE · CRITERION 3

preview-482 · /checkout

EXPIRED2025
Apply
Subtotal$84.00
Discount−$8.40
Total$75.60

Receipt

sha256:9f2c…e41a · commit a1b9c3e

Captured by the verifier in an isolated sandbox. Agent-written tests are not counted.

Illustrative example.

FITS THE STACK YOU RUN GitHubClaude CodeCodexCursorCopilotDevinVercelNetlifyLinear · soon

01 · THE PROBLEM

Writing code became cheap. Trusting it did not.

Review is the new backlog

Agents open pull requests faster than people can read them, and senior engineers turn into full-time diff readers.

Passing tests are not proof

When the same agent writes the code and the tests, both inherit the same wrong assumption.

Done was never agreed

A one-line issue leaves edge cases to the agent’s judgment. You find out what it decided after the merge.

Escapes repeat

Last month’s production bug rarely becomes this month’s check, so the same class of mistake ships again.

02 · HOW IT WORKS

One gate between your agents and your main branch.

  1. STEP 1

    Agree on done

    Label an issue. MergeProof drafts observable criteria from your code, schemas and past bugs. A person edits and approves them.

    1  Valid code takes 10% off · Browser
    2  Expired code is rejected · API
    3  Copy fits the brand · Human
  2. STEP 2

    Any agent builds

    Keep the agents you already use. They receive the approved plan and open a pull request with a preview deploy, as they do today.

    feat: discount codes at checkout
    7 files changed · preview-482 ready
    opened by claude-code
  3. STEP 3

    Verified, then merged

    An independent verifier exercises the live app and APIs, posts evidence on the PR, and holds a required check until every criterion is met.

    Proven  4 criteria with evidence
    Failed  sent back to the agent
    Sign-off  1 awaiting owner

03 · WHY IT’S DIFFERENT

Review reads the change. MergeProof proves the behavior.

CapabilityAI code reviewCI test suiteMergeProof
Checks against intent a person approvedNoOnly if someone wrote the testYes
Exercises the running applicationNo, reads the diffPartlyYes, on the preview
Independent of the authorYesNo, when the agent writes testsYes, separate sandbox
Evidence a reviewer can inspectCommentsPass or fail logsScreenshots, traces, signed receipt
Learns from escaped bugsNoManualEach escape becomes a criterion

MergeProof works alongside review tools and CI. It doesn’t replace them.

04 · SECURITY AND GOVERNANCE

Built to run untrusted code safely.

Verification means executing changes nobody has reviewed yet. Isolation and provenance are the product, not an add-on.

Isolated runs

Every verification gets a fresh sandbox, destroyed when it finishes.

Scoped credentials

Short-lived secrets per run; network egress limited to your preview.

Signed evidence

Each report is hashed and signed against the commit. A new push invalidates it.

Full audit trail

Who approved which criteria, what ran, and who signed off.

SSO and roles · roadmap

SAML single sign-on and role-based approval policies.

Self-hosted runners · roadmap

Keep code and data inside your own cloud account.

05 · PRICING

Start as a design partner.

Open source

Free

Public repositories with a monthly allowance of verified pull requests.

Join the waitlist

Team

Free for design partners

Private repositories, required merge check, sign-offs, and an included pool of verified pull requests. Pricing set with our first partners.

Apply as a design partner

Enterprise

Custom

Volume pricing, SSO, approval policies, self-hosted runners and a dedicated contact as they ship.

Talk to us

06 · QUESTIONS

What teams ask first.

Do you write code?

No. You keep your agents. MergeProof defines done and verifies the result, so it stays neutral across vendors.

What can it verify today?

Web applications and their APIs on per-PR preview deployments. Data migrations, queues and background services are next.

What if a criterion can’t be checked automatically?

It’s marked for human sign-off when the definition is drafted, not discovered at review time.

Which models does it use?

MergeProof is model-agnostic. Drafting and verification can run on the provider your organization approves.

Put a proof gate in front of main.

We’re working with a small number of teams that ship with coding agents every day. Design partners get the full product free.

We’ll only use your email to contact you about MergeProof.