Skip to content
CrossCutEspañolLog inSign up

Your AI assistant forgets.CrossCut remembers.

Review findings become lessons your team approves. Your assistant gets them before it writes code.

Illustrative example

Developer task: Add an endpoint that receives a callback URL.

Without CrossCut

app.post('/callback', async (req, res) => {

await fetch(req.body.url)

res.sendStatus(204)

})

Review comment

req.body.url isn't validated. Same mistake as last week.

With CrossCut

[RELEVANT LESSONS — do not repeat these mistakes]1. [a41f09c2] (area: api.ts) When building URLs from input data, check the protocol first: only http and https.[END LESSONS]

app.post('/callback', async (req, res) => {

const url = new URL(req.body.url)

if (!['http:', 'https:'].includes(url.protocol)) return res.sendStatus(400)

await fetch(url)

res.sendStatus(204)

})

Works with

  • Claude Code

You fix the bug. Your assistant never finds out.

The same bug, three reviewsIllustrative example
  1. You

    api/payments.ts

    await fetch(req.body.url)

    You fix it in review.

    Fixed

  2. A teammate

    api/webhooks.ts

    await fetch(req.body.url)

    The same bug, another review.

    Still there

  3. You, in another session

    jobs/reminders.ts

    await fetch(req.body.url)

    The same bug, another review.

    Still there

The fix stays in the code. Not in the assistant.

Other tools show up once it's already written.

  1. You ask for code
  2. CrossCut warns it about the bug

    The assistant writes
  3. It gets reviewed

    Human review, PR bots and SonarQube catch it

CrossCut remembers what they found

How CrossCut works

From a review comment to a rule that blocks the bug.

Capture

Catches the bug where you already review.

  • Start by asking your assistant for a code review: CrossCut saves the findings and you approve the lessons. GitHub, SonarCloud and SARIF are connected with the team's help.
  • Secrets and paths are removed before anything leaves your machine.

Finding

The URL is built from input data without checking the protocol.

Source
PR bot comment
File
src/api.ts
Status
Sanitized

Illustrative example

Lesson

CCM 2.0, CrossCut's own model, writes the lesson.

  • What happened, why, and how to avoid it, from the finding and its fix.
  • It stays a draft. Nothing is approved on its own.

Lesson Draft

What happened
A request was built from a URL that came from the user.
Cause
The protocol was not checked before using it.
Fix
Accept only http and https, and encode special characters.
Pattern
new URL(input) without checking url.protocol

Illustrative example

Approval

Your team decides what counts.

  • Approve, edit or discard from the inbox, with keyboard shortcuts.
  • If nobody reviews it in 90 days, it expires. You can bring it back.

Lesson

Accept only http and https, and encode special characters.

ApproveEditDiscard

draft, approved, discarded or expired. Today: draft.

Illustrative example

Reminder

The assistant gets it before it writes.

  • Only approved lessons arrive: the 3 closest to what you’re writing.
  • Today it works with Claude Code. Cursor, Grok and OpenCode are in testing.
[RELEVANT LESSONS — do not repeat these mistakes]1. [a41f09c2] (area: api.ts) When building URLs from input data, check the protocol first: only http and https.2. [7be2d310] (area: panel-config) Every new Vite proxy route needs its location block in nginx.conf.[END LESSONS]

Illustrative example

Rule

Whatever can be automated gets blocked on its own.

  • Each rule is tested on cases that must fail and cases that must pass.

Rule url-without-protocol-check

Active: blocks

  • Must fail: new URL(input) with no protocol check.
  • Must pass: code that checks http and https.

Illustrative example

CCM 2.0, the model that writes your lessons

It is CrossCut's own model. It reads each finding and how it was fixed, and writes the draft of the lesson: what happened, why, and how to avoid it.

Lessons with a valid format

185 out of 186 in our tests

100 out of 100 on real code-review cases from open projects

Is it a real mistake or a false positive?

Right in 240 of 264 cases from our tests

Speed

About 20 seconds per lesson; up to nearly a minute under load.

Your data

It was not trained on customer data.

What still goes wrong

Sometimes it flips a rule or misses a minor detail. That is why no lesson is approved on its own: a person on your team always reviews it.

Measured between November 9 and 11, 2026.

Integrations that exist today

Live is called live. A sample config is called a sample.

Findings coming in

  • GitHub App

    Live

    Inline comments and bot review summaries. We set it up with you.

  • PR review bots

    Live

    Works with any bot, CodeRabbit included.

  • SonarCloud

    Live

    Webhook and project connector, with an encrypted token. We set it up with you.

  • SARIF

    Live

    Sent through the API.

  • Review by your assistant

    Live

    Through the MCP tool capture_findings or a hook.

Assistants that receive lessons

  • Claude Code

    Plugin

    A hook on every prompt; a gate on Write and Edit.

  • Grok

    In testing

    Beta access by invitation.

  • OpenCode

    In testing

    Beta access by invitation.

  • Cursor

    In testing

    Beta access by invitation.

CrossCut

Unverified or unpublished

  • Codex, Gemini CLI and Continue

    Example

    Sample MCP configuration only. Not verified in production.

  • GitHub Action crosscut-gate

    Unpublished

    It exists, but it isn't published yet.

What we measured, and what we did not

We also tell you what still falls short.

Does it block the mistakes it already knows?

Yes, in this test.

17 files with a mistake: all blocked

17 clean files: none blocked

Does it catch the mistakes from your reviews?

Yes, in this test.

  • It caught the 18 expected findings

  • It made none up (15 test cases)

Does it find the right lesson?

9 times out of 10 on our test set

Measured with test lessons

With 1,000 lessons, almost 9 out of 10

These are test lessons, not customer lessons: the real measurement will come with real projects.

Does it stop the assistant from repeating mistakes?

From 60 down to 13 repeated mistakes.

Promising, internal test

  • Without CrossCut: it repeated the mistake 60 times out of 60

  • With CrossCut: it repeated it 13 times out of 60

160 runs with a real assistant. In 6 tasks that test a known mistake. In the other 2 tasks it did not repeat it either way.

Lessons written by our team and a single assistant: this is not yet a promise for your project.

More than 350 lessons approved by the team that builds CrossCut.

Technical detail
LayerWhat was measuredResultReadingNotes
Deterministic gate34 files: 17 with a mistake and 17 clean
  • Recall 1.00
  • Precision 1.00
  • False denials 0.00
Good on this corpusOn a live corpus of 300 cases, the 95% Wilson lower bound is 0.975.
Finding captureFindings collected from the sources
  • Recall 1.00
  • Precision 1.00
Good on this corpus
Lesson retrievalWhether the right lesson lands in the top 3 (60 queries, measured on November 1, 2026)
  • Recall@3 0.933 (300 lessons)
  • Recall@3 0.883 (1,000 lessons)
Measured with test lessons
Test with an agent160 runs with a real assistant, with and without CrossCut (internal test, November 6, 2026)
  • Without CrossCut 1.000 (60 runs)
  • With CrossCut 0.217 (60 runs)
Promising, internal test

Each layer has its own corpus. Don't compare across layers.

We haven't measured this with customers yet: we don't promise a specific reduction in your project.

“Mistakes avoided” in the dashboard is an estimate: times delivered minus repeated mistakes. It doesn't show whether the assistant used the lesson.

Your data, under your control.

  • Project A
  • Project B
  • Project C

Every project is a silo.

One project's lessons never mix with another's.

  1. We don't train on your data without your permission.

    Only projects that opt in. You can revoke it, and it's audited.

  2. Secrets and paths are removed before they leave your machine.

    The server cleans them again, emails included.

  3. We don't send your prompts to an AI provider to find lessons.

    The text of your prompts is deleted after 90 days by default. To set a different period, write to us.

  4. Take your data with you, or delete it, whenever you ask.

    Ask by email, at the contact in the privacy policy, to export your data, delete a project or delete your account, whatever your plan; we reply within five business days. You download the audit yourself from the panel.

No lesson approves itself. A person always reviews it.

What we don't have yet

  • We don't have SOC 2 yet. It's on the roadmap.
  • The legal texts are drafts, pending legal review.
Read the privacy policy

The providers that handle data will be listed in the privacy policy once it is published.

Frequently asked questions

Is CrossCut a code review tool?

No. Review stays yours, whether done by people, PR bots or SonarQube. CrossCut takes what the review found and gives it to your assistant before it writes new code. It's not an IDE and doesn't replace your assistant.

Does my code leave my network?

Only sanitized findings leave: the client removes secrets and paths, and the server sanitizes again, emails included. Prompts are processed on CrossCut's infrastructure; to find lessons they are not sent to any AI provider. The providers will be listed in the privacy policy once it is published.

Does it reduce repeated mistakes?

We haven't measured this with customers yet, and we don't promise a specific reduction in your project. In an internal test on November 6, 2026, with 160 runs and a single real assistant, in 6 tasks that test a known mistake it repeated the mistake 60 times out of 60 without CrossCut and 13 times out of 60 with it; in the other 2 tasks it did not repeat it either way. Our own team wrote the lessons. Details are in the evidence section.

Who approves the lessons?

Any project member can approve, edit or discard. The lead manages members, providers and integrations, and can require double approval for rules.

What does “avoided (estimated)” mean?

It's an estimate: times delivered minus repeated mistakes. It doesn't prove the assistant used the lesson. “Delivered” only means it was in a run’s block.

Which assistants does it work with?

Today it works with Claude Code. Cursor, Grok and OpenCode are in testing. For Codex, Gemini CLI and Continue we have sample MCP configurations, not verified in production.

How much does it cost?

Prices are decided at launch.

Can I sign up and start today?

Yes. Create your account with your email and sign in with a code we send you.

Don’t let the same mistake happen again

Create your account and get started today.

Sign up