English audit for Codex
You already write English in Codex every day. The audit reads that history back, finds your mistakes, and shows you the list — you review every line before anything goes to Glite.
Quick start
Codex asks for your consent before reading anything. Requires Python 3.12+ and uv (Codex installs it if missing). Pre-release: direct submission is off, the review page is download-only — upload the package here. Apache 2.0.
Who sees what
Three parties, three very different views. This is the whole privacy story.
- Yes: Your messages and full history
- Yes: All audit results, in one local folder
- Yes: Delete it anytime
- Yes: Re-reads only what you already typed into it
- No: Never another agent's history
- No: No file paths or project names
- Yes: The mistake list — rule + one safe example each
- Yes: Only after you review and confirm
- No: No messages, no names, no account
How it works
Paste one instruction
Codex clones the repo, installs what it needs, and reads the audit skill itself. It tells you what it will read and asks once for permission.
It reads and finds
Codex reads your own message history, picks a period (~2 weeks of ordinary use), and flags only what strongly signals non-native English — precision over recall.
You review, then send
A local review page shows every record. Include or exclude each one, then download the package — the list, not your messages — and turn it into your report.
This is all Glite ever receives
Real records from a real audit — each is just an example, what went wrong, and the rule.
to understand is it genially missing
Mistake: Formed an embedded question with direct-question word order and no 'whether'.
Rule: A yes/no question embedded in a sentence takes 'whether' or 'if' and statement word order: 'to understand whether it is missing'.
Why the total is 0 in the summary?
Mistake: Formed a direct question with statement word order.
Rule: A direct wh-question puts the verb 'be' before the subject: 'why is the total'.
“Verbatim” = your exact words, safe to send. “Redacted” = an identifying detail was replaced before you review it.
The questions people ask first
Does Glite see my messages?
No. It receives a list of mistakes — the wrong construction, the rule, and a short safe example. You read every line before sending and can exclude any of them. Excluded records send no details, only an anonymous +1 to the withheld count.
Is letting my agent read this a new privacy risk?
No new one. These are messages you typed into that agent; it received them when you wrote them. Reading them back discloses them to nobody who didn't already have them. The code enforces the boundary: the audit reads the history of the agent it runs inside, never another provider's.
Is it 100% local?
Honestly, no — and we don't claim it. Selected text is processed by the AI provider behind your own session, the same provider that already had it. This package itself makes no inference calls, no network requests during discovery, and Glite gets raw text never.
Can I delete what I sent?
No — because we can't know it was you who sent it. Submission is anonymous by design: no account, no email, no identifier. With nothing tying a package to a person, a deletion request can't be honoured — so sending is permanent, requires an explicit 18+ confirmation, and that's exactly why the review step exists: nothing is sent until you've read it. Today direct sending is off entirely; the review page is download-only.
Not using Codex?
Two other ways to get the same report.
Claude Code
The same audit, reading your Claude Code history instead. Identical privacy model.
Paste your text
No install, no history. Paste writing you already have and get the report — Glite analyses it on the server.
Already ran the audit and downloaded the package? Upload your results to get the report.