Get AEI Scan
To get AEI Scan, write to info@prosights.com.
What you get
A run writes 1 report in 3 parts.
Page 1, for a leader. Your result in a line, the level in plain words, the first 5 things to do and what the result rests on.
The 16 dimensions, each with a level and a score, in a single table.
An appendix with the detail. Every action, the file or setting that closes each open check, what passed, where your answers and the scan disagree, and every location read.
The report arrives as report.html, which opens in a browser with no connection, as report.md, and as data in results.json.
Read our own report: Kyberna, the product Prosights builds, scanned on 2026-11-16.
When a repository holds files of more than 1 coding agent, a run assesses each of them. Each further agent gets its own results file, and the report covers them all. With --benchmark the appendix gains a part, "How you compare".
Run it in 3 steps
AEI Scan needs Python 3 and nothing else. Nothing is installed.
1. See what it would read. A trial run lists every location and stops. Nothing is read.
python3 assess.py /path/to/your/repository --dry-run
2. Run it and answer the questions. Keep your answers for the next run.
python3 assess.py /path/to/your/repository --interactive --save-answers answers.json --out results
You are asked 3 opening questions, 18 survey questions and 35 questions on what your team believes is in place. Press Enter to skip any of them.
3. Read the report. A run on a made-up repository that uses Claude Code and Cursor printed these lines for Claude Code:
Platform: Claude Code
Result: AEI 84, Level 1 (provisional)
Overall level: AEI Level 1, Written (provisional)
Overall score: 83.8
Coverage: 53 of 97 checks (55%)
Claim gap: 5
Cursor, the second agent in that repository, got the same 6 lines with its own result. A level is provisional when at least 1 dimension had too few graded checks to be rated. The claim gap counts the checks where your answer was yes and the scan found nothing.
What it reads
AEI Scan reads only the locations that the vendor of each coding agent documents.
Whether each documented file exists.
In the agent's settings files, the keys that the checks name: the sandbox, the network list, the permission rules and the hooks.
The text of the agent's instruction files, to find headings, command lines and a few words, and to count lines.
The names of the files in
.git/hooks/and.github/workflows/, and on macOS and Linux whether each git hook file is executable.How many files the agent's memory folder holds.
How much the scan reads depends on your coding agent, because vendors document different amounts: 35 of the 35 scan checks at most and 3 at least. Where a vendor documents no location, AEI Scan does not guess, and a person answers the check.
| Coding agent | Scan checks the tool reads | Scan checks a person grades |
|---|---|---|
| Claude Code | 35 | 0 |
| Codex CLI | 22 | 13 |
| Cursor | 21 | 14 |
| Antigravity | 10 | 25 |
| Grok Build | 3 | 32 |
| VS Code with Copilot | 19 | 16 |
| Xcode | 3 | 32 |
The counts are for the system where the vendor documents the most. On Windows the scan reads 33 checks on Claude Code, and on macOS and Windows 18 on VS Code with Copilot.
What it never reads
It never opens a source file, a memory file or a session record.
It never copies the value of a setting, the command of a hook or a line of an instruction file into the report.
It counts session records only when you pass
--count-sessions, and never opens them.The report holds no user name, host name or home path unless you pass
--include-machine-identity.
A test backs each of these statements. The briefing for the technical team (PDF) names every test.
Nothing leaves your machine unless you say so
AEI Scan sends nothing by default. The last line of a run says so: "Nothing was sent anywhere."
The option --upload sends the results file to the results site. Before it sends, AEI Scan prints the size of the file, the address and what the file holds. No source file, answers file or report is sent.
It also prints the SHA-256 digest of the file, and the results site shows the same digest with the saved result. Equal digests mean the saved result is your file, unchanged.
The option --benchmark reads the published benchmark figures from the results site. The request holds no part of your result, and the address is printed before the request is made.
Keep a result to read later
The results site keeps the results you upload. You need no account to upload.
python3 assess.py --upload-file results/results.json
The site checks the file and recomputes the score from the check results. It refuses a file whose score does not match.
The site cannot check that the answers in the file are true. The report on the site says so.
You receive a claim code and a claim link. The code works once, within 7 days.
Open the link, then register or sign in. The result is saved to your account.
From then on, only your account can read or delete the result.
A result that nobody claims within 7 days is deleted.
The results site also takes the file through a form in your browser. What the results site stores lists every item the site keeps.
Answer an assessment request
An investor or a buyer may send your company a request link and a request code. The page behind the link says who asks, what would be shared and until when.
python3 assess.py /path/to/repository-1 /path/to/repository-2 --answers answers.json --request CODE --upload
Each repository you name is assessed, and its files go into their own folder.
Before anything is sent, you see what is sent: the results file, the request code and the name of the repository's folder.
Uploading needs no account. The requester reads nothing until a person at your company signs in on the results site and agrees.
In that account you read your results, remove an upload, and state which repositories are in the results and which are left out. Then you agree.
You can withdraw at any time. The site then deletes what it holds for the request.
The Benchmarks page describes the report that the requester reads.
Compare your result
python3 assess.py --benchmark-file results/results.json
Your overall score and your score in each dimension appear beside 3 things: a scan of public open-source repositories, a new installation of your coding agent, and the results that members of the results site chose to share. The Benchmarks page says what each of the 3 is and what it cannot show.
The published figures are read with a single request, which holds no part of your result.
A comparison is shown only for a group of at least 11 results. For a smaller group you see the count.
To share your own result, upload it, claim it and tick the box on the results site. The box is empty until you tick it.
What AEI Scan does not do
The scan shows that a control is present. It does not run a hook or test that a control works.
A coding agent is assessed only when the repository holds a file of that agent. An agent that is only installed on the machine is named and not assessed.
3 checks match headings, command lines and a few words in an instruction file. A file that says the same in other words fails those checks.
No study read for the index shows that a higher level brings faster delivery or better code.
Requirements, and how it was tested
AEI Scan is a Python 3 program that imports only the standard library. Prosights wrote and tested it on Python 3.9.6 on macOS. The tests give it the folders, environment variables and file formats that each vendor documents for Linux and Windows. On 2026-11-14 the test suite passed on GitHub's Linux and Windows machines; 3 tests that need a Linux or macOS file system are skipped on Windows. No person has run the program on a Linux or Windows workstation with real agent files.
The Prosights Agentic Engineering Index page explains the levels, the dimensions and the score.
Start with the trial run
Write to info@prosights.com for AEI Scan, then run step 1. The trial run lists what would be read, and reads nothing.
Prosights.ai