See what they could do before you click Add.

The smallest access that works.

Read-only: nothing changes.

HOW IT WORKS

Two parts.
Only one uses AI.

  1. Pick a repo
  2. Pick a role
  3. See what it can reach

1. Code checks the repository. Ajar reads the repository’s settings on GitHub (permissions, workflows, approvals, protections) and works out what each role can reach. Fixed rules, no AI. That is the house scene and the access report.

2. AI reads only the access request. When someone asks for access, AI reads the request and points out pressure patterns seen in past takeovers. It never analyzes the repository, and it talks about the request, never the person.

3. Three AIs vote separately. Claude Opus 5.5, GPT-5.6 Sol and Gemini 4 Argon each read the request alone. The code counts the votes: a level or a signal shows only when at least 2 agree. Otherwise you see “no agreement”.

4. Here, or in your terminal. This site is the live demo for public repositories. In your terminal, ajar check OWNER/REPO runs on your own repositories, private ones included, with your own GitHub login.

Reading the scene

The three characters stand in for one person with the role you chose. They are guides, not AI agents, and nothing runs against the repository. The AI results on this site are pre-computed; no AI is called here.

Experimental: file check

The same 3 AIs can also be asked which files in a change deserve a careful look. In a small blind test with 8 files (3 with a planted problem, 5 normal), the panel flagged 3 of 3 and 0 of 5. Alone, Claude declined 2 files and Gemini raised 1 false alarm; a declined file still counts as needing review. A test this small is not a general measure of accuracy. Not part of this site.

EXAMPLE REQUESTS

The request matters, too.

Look at the wording, not the person.

The maintainer decides.