Skip to content
For teams shipping code written by AI agents

Your agent writes it. MLReps checks it.

MLReps reads your whole repository, draws every route from the open internet to your data, and checks every pull request against main before you merge.

We’ll email you a sign-in link. A password is optional.
$3 freeNo cardmacOS, Linux and Windows
Watch a pull request get checked
MLReps › acme/checkout › PR #482 report
Ready when you are.

An example, not a measurement — this repository is made up.

Measured, not claimed

See a real report. No account needed.

Six open-source projects, read at a pinned commit on 8 November 2026. These are the engine’s own recorded answers — the figures npm run golden holds it to on every release.

Pick a repository
Recorded 8 November 2026 · pinned commit
dubinc/dub
No report published yet
4,368
files read
16,847
nodes in the graph
305
findings
99%
of calls resolved
Commit 7a7beca. Security findings on a public report are counts only — no file, no line, no detail.

No report has been published on this deployment yet, so the numbers above are the measurement and not a page you can open. They are the same figures a published report would carry.

Set it up on your project

One command on your machine. Or three clicks on GitHub.

~/acme/checkout
$ npm install -g mlreps
$ mlreps setup
Open the link in your browser and approve code QF7K-2MPD
✓ Signed in on this device
? Run MLReps on every pull request? (Y/n) y
✓ Added .github/workflows/mlreps.yml — posts “MLReps · CI” on every PR
Reading ./checkout on this machine. Source never leaves it.
1,284 files read · 6 could not be read
⚠ 1 exposed path · 1 broken connection · 1 slow query
A local run reports counts by kind and the top findings. The full findings list, the architecture and what a change reaches are on the report this publishes to.
TypeScript is built in: the CLI reads it with nothing installed. JavaScript is built in: the CLI reads it with nothing installed.
Live today:npm install -g mlrepsnpx mlreps@latest scan .pnpm add -g mlrepsyarn global add mlreps
Node.js 20 or newer on your machine, or a build that needs no Node at all from the install script or Homebrew. Teams without Node use the GitHub App instead, which installs nothing.
No Windows build yet: the binaries here are for macOS and Linux, x64 or arm64. The npm package installs wherever Node 20 runs, Windows included.
1Device login. Approve a code in your browser.
2Scan on your machine. Your source stays there.
3The top 3 in your terminal. With counts by kind, free.
4Publish for $1. Every finding, the map, and what a change reaches.
Best together
Before you pushThe CLI catches it on your machine — one command, before you open the pull request.
Before you mergeThe “MLReps” check catches it on every pull request — from the GitHub App with no setup, or from your own CI as “MLReps · CI”, where your source never leaves your pipeline.
The GitHub App
Pull-request checks with no setup — installs in three clicks.
1
Install the App
ContentsReadMetadataReadPull requestsReadChecksRead & write
2
Choose repositories
acme/checkoutfirst audit of main
3
Open a pull request
Add upload endpoint for customer documents #482
4
The check appears
✅ No new issues vs main⚠ 1 new exposed path vs main
1PR checks

Merge like you read every line.

MLReps posts a check on every pull request, compared against main. Only what this pull request changes counts.

No new issues — merge.
New issue — look before you merge.
New critical — fix it first.
Add upload endpoint for customer documents #482
Openfeat/avatar-upload → main
src/lib/fs.ts
17export async function writeFile(userPath, data) {
18 const target = join(UPLOAD_DIR, userPath)
19 await fs.writeFile(target, data)
path traversaluserPath comes straight from POST /api/upload, a public route. A request can write outside the uploads folder.
MLReps
⚠ 1 new exposed path vs main
See it on the map →
InternetPOST /api/uploaduploadHandler()writeFile(userPath).env · user files

An example, not a measurement — this repository is made up.

MLReps › Reports › acme/checkout · main
Audit report · main
acme/checkout
1,284 files read · 6 could not be read
Share linkDownload
One route from the internet can write anywhere on disk. One call points at a function that was removed. One request repeats the same query 1,240 times.
Reachable from the internet · 1
A public upload route can write anywhere on disk
Path traversal · POST /api/upload → src/lib/fs.ts:18
See it on the map →
Broken connections · 1
Deleting an account calls a function that no longer exists
Broken connection · DELETE /api/account → purgeUser() · removed in c2a1f7
See it on the map →
Slow · 1
One request runs the same query 1,240 times
Query inside a loop · src/db/customer.ts:14
See it on the map →

An example, not a measurement — this repository is made up.

2Audit reports

A report, not a dashboard.

One page per audit, in plain words first and file:line second. Send the link to your team or a client.

Ordered by what the internet can reach
Every finding opens on the map
Files we could not read are listed, with why
3Optimization reports

Find the slow parts before your users do.

For the code that works but costs you: queries inside loops, N+1 queries and heavy files — each with the line to change.

Query inside a loopN+1 queriesHeavy files
Optimization report · acme/checkout · main
One request runs the same query 1,240 times
Query inside a loop · src/db/customer.ts:14
Slow
13for (const id of ids) {
14 const row = await db.query(SELECT_CUSTOMER, [id])
15 rows.push(row)
Suggestion: load them in one query, outside the loop.See it on the map →

An example, not a measurement — this repository is made up.

Findings · acme/checkout
Ordered by what the internet can reach, then by what is new. All 3.
NEW vs mainPath traversalReachable from the internet
A public upload route can write anywhere on disk
POST /api/upload → src/lib/fs.ts:18
Broken connectionReachable from the internet
Deleting an account calls a function that no longer exists
DELETE /api/account → purgeUser() · removed in c2a1f7
Credentials in codeNot reachable from the internet
Credentials are written in the code
src/config/stripe.ts:7
Credential-shaped, not confirmed as live — and the value itself is never read or stored.

An example, not a measurement — this repository is made up.

4Findings

Every way in, in order.

The loopholes someone could use: an upload route that writes anywhere, credentials left in code, calls to functions that are gone.

No scores and no grades — just what is reachable, what is new, and where to fix it.
How it all connects

One map behind every check, report and finding.

You see it the first time, and again when something breaks and you go looking.

From the open internet to your files and .env, through a public upload route.
Hover or tab to a station to see its file, line and code.

An example, not a measurement — this repository is made up.

Measured, not claimed

Run on 6 open-source repositories.

Pinned commits, recorded 8 November 2026. These are the engine’s recorded answers, held on every release by npm run golden.

dubinc/dub
4,368
files read
16,847
nodes in the graph
305
findings
99%
of calls resolved
7a7becaNo report published yet
outline/outline
2,518
files read
12,143
nodes in the graph
161
findings
96%
of calls resolved
0ed4a80No report published yet
documenso/documenso
2,112
files read
9,632
nodes in the graph
793
findings
100%
of calls resolved
da934c0No report published yet
fastify/fastify
297
files read
1,098
nodes in the graph
1
finding
44%
of calls resolved
630acd0No report published yet
expressjs/express
141
files read
469
nodes in the graph
18
findings
97%
of calls resolved
9a34acfNo report published yet
pallets/flask
83
files read
1,547
nodes in the graph
3
findings
98%
of calls resolved
ab81496No report published yet

Why one of them reads 44%. Fastify wires itself up at runtime through plugins and decorators, which a reader that follows code rather than running it cannot trace. That is a limit of ours, not a verdict on their code — and it is on this page because a repository we read badly is the one you most need to see before you trust the others.

Pricing

Per repository. Never per developer.

Add the eleventh developer and the bill is the same.

Priced per seat
11 developers, 11 seats — $264 to $792 a month.
The bill grows with every person you add — and with everyone who only reads it.

CodeRabbit’s published plans, billed annually — their pricing page, read 17 November 2026. How we compare.

MLReps
acme/checkout+ another repo
One price per repository, however many of you there are.
Push forty times and it is the same charge. Ten more people on the team and it is the same charge. Nobody on your team has a seat, because there are no seats.
GitHub App
$1$1 per repository, per day it’s audited — pushes and pull requests share that day. Up to 10 audits a day included.
  • Audits the default branch of the repositories you pick
  • A check on every pull request, against main
  • Audit report, optimization report, findings and the map — no separate report fee
  • Runs on our worker inside a sandbox with no network, and the checkout is deleted when the analysis ends
CLI
Freeto scan, on every repository.
  • Your source never leaves your machine; the report is uploaded so you can open and share it.
  • The scan and its findings in your terminal are free, every time, on every repository
  • Publishing a report is what costs — the same commit published again costs nothing
Pull-request checks from your own CI, so source never leaves your pipeline.
Verify your email and you get $3 of balance — 3 published audits, with nothing to enter.
No card to start.
Never charged
  • Unchanged code
  • The local scan
  • An audit we could not do
  • A repository already audited today
  • Anything that failed on our side

What each of those means

Security & privacy

We read your code. We don’t keep it.

Never stored

Your source code is never stored.

No column exists that can hold source, a diff, a patch or a token — retaining one would take a database migration, and a build gate fails any log call that could receive source. On the command-line path your code never leaves your machine at all. On the GitHub App path it is unpacked onto an analysis worker, read inside a sandbox with no network, and deleted when that one analysis ends.

And not to a model

Nothing about your code is sent to an AI model — not your source, and not metadata about it.

No file path, no line number, no operation name, no count and no finding of ours leaves for a model, and no model writes a word you read. Three things make that so rather than one: the one screen that sent a model anything no longer does, no model key is configured in the service at all, and a build gate counts every place in the code that could reach one. If we ever turn on the assistant that would send a question and its evidence to be worded, /security says what that sends and this page will say it before it happens.

Kept, and for how long
  • Repository names — Until you uninstall, which removes them at once.
  • Findings and the architecture map — Until you ask us to delete it. Uninstalling the App does not erase it — see below.
  • Your account — Until the account is deleted.
  • A payment reference, if you pay — For as long as the balance exists. The ledger is append-only, so a credit is never edited or removed — a correction is another entry beside it.

One organisation can never read another's data.

GitHub permissions
ContentsReadMetadataReadPull requestsReadChecksRead & write

The one write is for posting the check on your pull request, and nothing else.

There are things that are not true yet — uninstalling does not erase your history, there is no self-serve delete or export, and operator access is not yet mechanically blocked. /security says all of it, including what we have not built.

What it refuses to say

The reason to believe the numbers is the ones it will not print.

  • A score it did not measure.

    No index, no score and no debt figure is printed anywhere — not a number on a scale, not a ranking, not an estimate in hours. What you get is counts, what was read, and what a change reaches.

  • A map of a codebase it could not read.

    Below 70% resolution no picture is drawn at all; between 70 and 90 it is drawn with its gaps named.

  • That a change is safe.

    An empty blast radius means we found nothing, never that nothing is affected.

  • Anything about a person.

    No per-developer score, no leaderboard, and no column anywhere for who dismissed a finding.

Check your next pull request.

We’ll email you a sign-in link. A password is optional.
$3 freeNo cardmacOS, Linux and Windows