AI (Claude / OpenAI / Gemini)
AI in AppArch is configured once for the whole installation, by the installation’s operator. Workspaces cannot add their own key, and there is no AI entry under Integrations.
Configuring it
Section titled “Configuring it”Open Administration → Settings → Assistant (/management/settings/assistant) and set:
- Provider - one of Anthropic (Claude), OpenAI or Google (Gemini). One provider at a time.
- API key - the operator’s own key; usage is billed by the provider to that account.
- Models on offer - every model this installation may use. The first one is the default. When more than one is listed, users pick per question in Archito’s Ask tab.
- Offer the assistant to users - turns on Archito’s Ask tab.
- Allow fix suggestions on findings - a separate switch, because each suggestion sends a finding to the provider.
- Token budgets - optional daily and weekly ceilings, per workspace and for the whole installation.
Everything is off by default. Neither the assistant nor fix suggestions can be switched on until the operator accepts the AI processing agreement on the same page.
The API key is encrypted with the installation’s master key. After saving it is never shown again - only its last four characters.
What uses it
Section titled “What uses it”| Feature | Turned on by | What is sent to the provider |
|---|---|---|
| Fix suggestion on ArchPilot and ArchScan findings | Operator: Allow fix suggestions | Rule key, severity, file path and line, the finding message and the analyzer’s own suggestion |
| Archito → Ask tab | Operator: Offer the assistant | The question and the figures used to answer it (project names, finding counts, scan totals) |
| AI verdicts on MR review comments | Workspace: ArchPilot settings, opt-in; needs either operator switch | The pending review comments being classified |
| MD generator, AI mode | Operator: Allow fix suggestions | The project facts, the rendered template draft and a shallow slice of the repository: description, topics, root listing, a few manifest files (e.g. package.json, *.csproj, go.mod, truncated) and the existing README |
None of them sends the repository as a whole. How the assistant looks to users: Archito, the assistant.
Fix suggestions
Section titled “Fix suggestions”- The answer is written in the reader’s panel language (Turkish, English, German, French, Spanish, Russian, Arabic, Azerbaijani; anything else falls back to English) with the default model.
- Every answer is stored with the finding, with provider, model, input and output token counts and language. The next person who opens the same finding sees the stored answer and no tokens are spent again. Regenerating it is a deliberate action.
MD generator
Section titled “MD generator”Without AI, the MD generator works in template mode (English only). See MD generator.
Limits
Section titled “Limits”- Rate limit: at most 30 AI calls per workspace per minute across all features. The counter lives in Redis; if Redis is unreachable the limit is not applied.
- Token budgets: per workspace per day and per week, and for the installation per day and per week.
0means no limit. The day resets at midnight UTC, the week on Monday. The installation ceiling always wins. Budgets count the Ask tab and review comment classification.
Network
Section titled “Network”The Api calls the provider directly:
https://api.anthropic.comhttps://api.openai.comhttps://generativelanguage.googleapis.com
There is no AppArch-hosted relay. On an isolated network, leave AI off - there is nothing for it to reach.