ai is the platform's single door to language models. No module calls a provider directly: every inference request goes through the gateway, which routes, meters, and audits it.

Current state

• RouterBackend routes requests across providers; local backends (LM Studio, Ollama) run without any cloud key, so the platform works fully offline. Cloud providers slot in behind the same interface when keys are configured.
• Cost controls: a cost ledger with reconcile, spend caps, and cost-hold expiry - a request that would breach the cap is refused, never silently billed.
• The embed sidecar serves embedding requests for the memory module's brain (deterministic stub in tests; real models in deployment).
• Everything is tenant-scoped and observable; gateway-only model access is a protected invariant no task may weaken.

Changelog

History lives in the changelog; this page describes only the current state.

Generated from services/ai-gateway/docs/index.md — edit the markdown source, not this file (TASK-DOCS-002).