> ## Documentation Index
> Fetch the complete documentation index at: https://docs.platform.aiplanet.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Capabilities and limitations

> What the gateway supports today, what it does not, and known accounting and deployment limits.

## Supported

* OpenAI-compatible chat completions.
* OpenAI-compatible Responses.
* Embeddings.
* Bounded OpenAI-compatible audio transcription for explicitly configured and live-validated STT aliases.
* Bounded OpenAI-compatible English audio translation for explicitly configured and live-validated OpenAI aliases.
* OpenAI-compatible chunked speech synthesis for explicitly configured OpenAI, ElevenLabs, or Azure Speech TTS aliases.
* Provider-native OpenAI Realtime and ElevenLabs realtime STT/TTS WebSocket sessions for explicitly configured `openai-realtime-v1`, `elevenlabs-realtime-stt-v1`, or `elevenlabs-realtime-tts-v1` aliases.
* Optional gateway-issued, short-lived, single-use browser credentials for gateway-terminated Realtime WebSockets.
* Optional provider-direct OpenAI WebRTC bootstrap with opaque single-use gateway credentials and exact-origin CORS. Media bypasses the gateway, so this path rejects budgeted keys and payload-retained deployments.
* Server-Sent Events streaming and disconnect finalization.
* Tools, vision, reasoning, and embeddings capability enforcement.
* Stable aliases, weighted routing, retries, fallbacks, and cooldowns.
* Private UI and API administration for versioned concrete models, weighted aliases, and scoped keys.
* Scoped hashed keys, expiry, revocation, RPM, TPM, and budgets.
* Tenant-owned admin-created keys and atomic token rotation with bounded prior-token overlap.
* Postgres usage/cost analytics and budget ledger.
* Redis-backed distributed limits and cooldowns; direct-HTTP cooldown state falls back locally during Redis routing-state outages without weakening fail-closed limits.
* Prometheus metrics, structured logs, readiness, and Slack or Google Chat budget alerts.
* Append-only administrative audit attempts/outcomes with fail-closed mutation preflight.
* OIDC Authorization Code + PKCE with immutable subject-to-role/tenant mappings and one-time shared login state.
* Optional AI Planet Platform SSO handoff with organization-membership role mapping, single-use tokens, and scoped/global session revocation.
* Automatic Postgres-backed model-catalog generation propagation with per-instance acknowledgements, component degradation, and last-safe-runtime serving.
* Payload persistence disabled by default and optionally enabled with an independent retention window, plus Postgres-leased metadata/audit and expired-auth-artifact retention.
* Optional Python trajectory SDK with nested runs, spans, tools, feedback, bounded OTLP export, privacy-safe payload defaults, and gateway generation correlation.

Provider and model support depends on the routes enabled and live-validated in your deployment. An alias appearing in a sample file is not proof that your credentials or provider endpoint support it.

Realtime support remains an internal pilot surface. Redis-backed admission and browser-credential consumption are shared across replicas; active sessions revalidate their parent key and renew session-second budget authority. ElevenLabs realtime TTS reserves the bounded maximum text characters and settles to accepted characters. Enable only aliases that have passed provider-backed close, outage, budget, shutdown, and representative-load drills in the target deployment.

## Not currently supported

* Anthropic-native Messages endpoint.
* Managed file uploads, batches, image generation, additional realtime agent protocols, assistants, or threads.
* MCP or A2A gateway endpoints.
* Automatic provider invoice reconciliation or invoice-grade billing.
* Dynamic IdP group provisioning, general key unblock, or scheduled rotation orchestration. OIDC subject mapping, local break-glass login, system/organization RBAC, atomic key update and rotation, and admin audit events are available on the private operator surface.
* Per-tenant alert destinations and policy plugins.
* Certified production capacity sizing or a recommended high-availability topology.

## Accounting limitations

Budget settlement uses provider cost when available, otherwise observed usage with reviewed pricing, otherwise the conservative reservation. This is designed to prevent budget bypass, not to reproduce a provider invoice exactly.

Request logs are asynchronous and can be dropped during a prolonged database outage. Budget admission uses a separate transactional ledger.

## Deployment limitations

The supported operating baseline remains a single gateway process until capacity and production HA procedures are established. Redis and Postgres provide multi-instance correctness for limits, budgets, login state, and model-catalog propagation. General deployment configuration reload is still process-local and must be rolled out consistently.