ManiLens

My own code reviewer. It reads every pull request in my projects, finds bugs and security holes, and writes directly in the code, in Danish, what is wrong and how to fix it. It is in development, and the case is updated along the way. The video was recorded with test data.

September–October 2026In developmentPythonClaude CodeGitHub ActionsPHP
ManiLensTest data

In short

21 of 31of CodeRabbit’s real bugs found on 10 PRs
about 90%of ManiLens’ own findings were real
15real bugs that CodeRabbit had not found
6steps in every review, in a fixed order

Why

I used CodeRabbit to read my pull requests before they were merged. I wanted the same thing, but in Danish, with my own rules for each project and without an extra subscription. ManiLens runs on the Claude subscription I already have.

The rules are what matter most. Every project has its own requirements, for example that a CHANGELOG must be updated, or that pages with personal data must not be indexed. A general reviewer doesn’t know them. ManiLens reads them from the project before it reads the code.

How it reads a PR

When a pull request is opened or gets a new commit, ManiLens runs the same steps every time:

A review6 steps · fixed order
  1. Fixed checks the project’s own tests, lint and typecheck

  2. Scanners up to 14, e.g. opengrep, Ruff, Gitleaks and Trivy, only where they fit the code

  3. Four reviewers one reads for bugs from the start and one from the end, one for security and one for the project’s rules

  4. Verifier rereads the code and drops everything it can’t confirm itself

  5. Verdict block the merge, comment or approve

  6. On the PR comments in the code with a fix and a ready prompt for an AI agent

A missing or failed check never counts as passed, and a block is never changed into an approval.

The verifier is why it can be trusted. The reviewers may suggest a lot, but only what is confirmed in the code reaches the PR.

A real comment

Here is a real finding from a demo project I use to test ManiLens. A search function put the user’s search term straight into the SQL.

Serious· Security · noter/lager.py line 41
The code
"… LIKE '%" + ord + "%' …"
  • The search term don't makes the command crash
  • x' OR '1'='1 returns all notes instead of the ones searched for
  • The two other functions in the same file do it right with ?
The suggestion
LIKE ? ESCAPE

You can ask about a finding directly in the PR, and ManiLens replies in the same thread:

Thread on the PRShortened

@manilens why is this a problem?

Reading noter/lager.py · done

Because the search term becomes SQL code instead of data. An apostrophe makes the command crash, and the right text can change what the query finds. It is serious, but not critical, because the only caller today is the local command line.

The answer is based on the code in the PR.

On GitHub and locally

ManiLens comes in two versions that use the same reviewers:

  1. 1On GitHubruns by itself on every pull request and writes the comments directly in the code
  2. 2Locally in Claude Codethe same reviewers as agents running on my own machine before I merge

That turned out to matter. At the start of October GitHub’s free minutes were used up, and reviews on GitHub stood still. The local version kept running, because it doesn’t use GitHub minutes.

Measured against CodeRabbit

I measured ManiLens on 10 real pull requests from my projects that CodeRabbit had already read. A third model decided whether each finding was a real bug.

ManiLens found 21 of CodeRabbit’s 31 real bugs. About 90% of ManiLens’ own findings were real, and CodeRabbit had not found 15 of them.

The bugs it missed were mostly in frontend code and in very large PRs. So it now gets large PRs in chunks and a frontend checklist. That was measured on 9 new PRs that were not used to write the checklist:

CodeRabbit bugs found (of 45)

12→17

Before
Now

Real bugs missed

11→6

Before
Now

The numbers come from a few runs, and the judging model doesn’t always agree with itself. They show the direction, not a guarantee.

How far it has come

Today ManiLens reads pull requests in my own projects. It has a server with login and an account page where you can see your reviews, and a small group of colleagues are trying it out.

How far it has come4 parts
  • Reviewbuilt and in use
  • Localbuilt and in use
  • Colleaguestesting under way
  • Fixes by itselfbuilt, but switched off

It can also fix the code itself, write tests and resolve merge conflicts, but that is switched off until it has been tested properly. Large PRs are still the weakest spot. The case is updated as it gets better.

Shall we talk?

I reply within 1–2 working days. I’m the one who replies, and the one who builds the site.