OpenAI

Fix suggestions from the GPT-5 family

Your operator connects an OpenAI API key once, for the whole installation. Every "Resolve with AI" request under an ArchPilot finding then goes to GPT-5, GPT-5 mini or GPT-4.1 from your own account - nothing passes through AppArch.

Auth: OpenAI API key, set by the installation's operator

Screens from a running installation

How it connects

01
Get a key
Create an API key on the OpenAI platform; usage is billed to your own OpenAI account.
02
Enter it once
In Management > Assistant the operator picks OpenAI and the models: GPT-5, GPT-5 mini or GPT-4.1. The first one chosen is the default.
03
Resolve a finding
Hit the button under an ArchPilot finding; the suggestion arrives in a code block, in the reader's language.

What you get

Paid for once

A suggestion is stored per finding, so the next teammate who opens it spends nothing.

Token receipts

Every suggestion records its input and output tokens, so the spend is auditable.

Rate-limited per workspace

One member cannot drain the team's credit in a loop; a stored answer never counts against the limit.

Operator's switch

Suggestions can be turned off without removing the key; answers already paid for stay readable.

How it works inside AppArch

One key for the whole installation, set by your operator. Calls go straight from your server to the provider, only when a feature asks for one.

What the key is used for

Fix suggestions on ArchPilot findings, stored per finding so the next reader pays nothing
Archito's Ask tab: it picks from a closed question catalogue and answers only about your workspace
Classifying human review comments on merge requests, twenty per call, if the workspace opts in
Writing Markdown documents such as README and CONTRIBUTING in the language you choose

Guards

At most 30 calls per workspace per minute
Optional daily and weekly token caps, per workspace and for the installation
Every call records its input and output tokens; AI Impact shows the spend per person and model

What it never does

Archito's learning never uses it; that model is trained on your server
Nothing passes through AppArch's own servers
Off and keyless until your operator turns it on

Credentials

One API key for the whole installation, entered by your operator under Management, stored encrypted with the installation's master key. The model list is enforced on the server.