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

# September

September brings two new provider expanded guardrail coverage including MCP tool calls, and a new lifecycle stage to inspect traffic before it ever reaches a model.

We've also shipped improvements across Anthropic, Azure, and Bedrock.

See what's new:

## Summary

| Area | Key highlights |
| :- | :- |
| **Gateway** | • Decisions API  • Per-attempt retry logging |
| **Guardrails** | • startHooks lifecycle stage  • MCP Gateway guardrails |
| **Models and providers** | • Meta  • ElevenLabs  • Workload Identity Federation  • xAI  • Claude Code |

## Gateway

### MCP Gateway Guardrails

MCP tool calls now go through guardrails before and after execution, using the same checks available for LLM requests. Denied calls return a structured error with the guardrail and check ID, so you know exactly what blocked a tool invocation.

Learn more: [MCP Gateway guardrails](https://portkey.ai/docs/product/mcp-gateway/guardrails)

### Decisions API

Typed-judgment models now have their own endpoint. Send a question with structured state and get back a typed judgment—boolean, choice, or score—with probabilities and a calibrated confidence score. Starting with **TypeSafe's Jev** family, these calls get the full gateway stack: authentication, routing, retries, caching, logging, and cost tracking.

Learn more: [Providers](https://portkey.ai/docs/integrations/llms)

### Per-Attempt Retry Logging

Every retry attempt now gets its own log row. A chain of failures was previously only visible as the final result—now each step is traceable with its own status and response.

Learn more: [Automatic retries](https://portkey.ai/docs/product/ai-gateway/automatic-retries)

## Guardrails

### New Lifecycle Stage: startHooks

Traffic is now screenable before authentication, routing, or any other hook runs. The new **startHooks** stage lets you inspect and normalize incoming requests at the very start—before the gateway commits to a provider.

Two new checks are available at this stage:

* **Header Check** — allow or block based on whether request headers contain specific key-value pairs
* **Header Transform** — remove, set, or rewrite headers before routing. Protected headers like `x-portkey-`\* can't be modified.

Existing checks—`allowedRequestTypes`, webhook, and JWT header lookup—can also now run as startHooks.

Learn more: [Guardrails](https://portkey.ai/docs/product/guardrails)

## Models and providers

### New providers

* **Meta** — Chat Completions, Messages API, Responses API, and Files API
* **ElevenLabs** — text-to-speech and speech-to-text, with full observability and reliability features

\*\* Other updates:\*\*

* xAI now supports image generation with OpenAI-compatible parameters plus xAI-specific aspect ratio and resolution controls.

* Claude Code and other Anthropic-native clients now discover models through `/v1/models`. The gateway detects Anthropic-protocol callers and returns Anthropic's native model shape instead of the OpenAI-normalized shape.

## Bug fixes and improvements

* JWT auth now accepts multiple JWKS URLs and attributes requests to the matching workspace user
* Conditional routing now reads fields from multipart/form-data requests (image edits, audio transcriptions)
* Messages API: reasoning models now stream their thinking phase as a thinking content block

## Support

> [card] [Need Help?](https://github.com/Portkey-AI/gateway/issues)
>
> Open an issue on GitHub

> [card] [Join Us](https://portkey.wiki/community)
>
> Get support in our Discord


## Related topics

- [September](https://docs.portkey.ai/docs/changelog/2025/september.md)
- [Embeddings](https://docs.portkey.ai/docs/integrations/llms/openai/embeddings.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.