AI keys, budgets and limits
Bring your own AI keys, choose which model does what, route cloud agents through Skaftor's gateway, and cap spend per work order, per QA run and per month.
Skaftor runs on your organisation's own AI keys — nothing is generated without one. Admins manage them in Settings › "AI keys & limits".
Provider keys
Add keys for Anthropic (Claude Code), OpenAI (Codex), Google Gemini and DeepSeek with "Add key", then "Save & verify". You can also add up to three OpenAI-compatible endpoints for AI features. Anthropic is the default.
Which AI does what
Use a different model for some tasks — main tasks vs. other AI features — and set a default effort.
The model gateway
With "Cloud agents go through Skaftor" on (recommended), cloud runs reach the provider through Skaftor. Skaftor holds the key, meters real usage, limits which models a run may call, and refuses the next request once a budget is reached.
Budgets and limits
- Monthly budget — one cap across providers, or one per provider.
- Work orders — most per feature, how many at once, time and money for each, and what happens if one runs out of time.
- Pull requests — whether CI is required and whether AI checks the code against the PRD.
- QA — whether the QA agent runs after every deploy, its per-run budget, and what happens when automatic QA fails.
- Where work runs — anywhere, or only in your own cloud.