Weekly Update — July 6 – 13, 2026

[missBerg]

A steady week of shipping: per-request backend credentials landed, Anthropic streaming token accounting got a thorough cleanup, the Helm chart picked up production-hardening knobs, and MCP saw both efficiency and protocol-conformance work. Thanks to everyone who shipped, reviewed, and filed. 🎉

🛣️ MCP spec RC prep, continued

Work toward the MCP spec Release Candidate (July 28) is moving from planning into structure. From the maintainer call: the conformance gaps in #2323 are being clustered by severity and broken out into GitHub sub-issues, with a companion doc narrowing the spec changes down to what specifically needs to change in AI Gateway. If you build on the MCP proxy, the sub-issues are where the pick-up-able work will appear.

The call also revisited the required-header proposal (#2309) — the current direction is to check whether the existing MCP security policy CRD with a CEL expression covers the use case before adding a new field to the forwarding spec.

✨ What's new

Auth & security

• Per-request backend credentials — #2253 by @kanurag94. Backend credentials can now come from request headers or dynamic metadata instead of a single static secret per backend — the building block for per-tenant and per-user upstream auth.

Provider & translation coverage

Token accounting was the theme of the week — three fixes that make usage and cost metrics trustworthy:

• Complete usage on Anthropic streaming responses — #2292 by @pjdurden. Input and cache token usage reported on message_delta events is now merged into the final tally, so streamed Anthropic responses report full usage instead of output tokens only.
• No more double-counted cache tokens — #2236 by @Aias00. Anthropic streaming cache tokens were being counted twice in metrics; operators tracking spend now get the real number.
• reasoning_effort → Bedrock reasoning_config — #2190 by @Killusions. OpenAI-style reasoning_effort now translates to AWS Bedrock's reasoning_config, so reasoning models behind Bedrock honor the client's effort setting.

MCP

• One Backend per namespace — #2257 by @kanurag94. The MCP proxy now shares a single Backend resource per namespace instead of creating one per MCPRoute — less resource sprawl for clusters running many routes.
• Tolerant initialize over SSE — #2267 by @pjdurden. Non-response SSE events during the initialize handshake are now skipped, so backends that emit pings or notifications mid-handshake no longer break session setup. Closes #2219.

Developer experience

• Helm chart production hardening — #2335, #2336, #2337, and #2338 by @walsm232. PodDisruptionBudget, topologySpreadConstraints, custom pod labels, and security-context defaults for the controller — the knobs operators expect when running it as a production workload.

Bug fixes & stability

• Cross-namespace InferencePool routing — #2339 by @Aias00. The namespace on a cross-namespace InferencePool backendRef is now honored instead of falling back to the route's own namespace.

💡 New design proposals & feature requests

One new thread worth a read:

• Compressed filter config — #2341 by @xiaolin593. Proposes gzip-compressing the extproc filter config so deployments with many model routes stop hitting the 1MB Kubernetes Secret limit — in testing, 1.3MB of config compresses to 79KB.

👀 PRs looking for review

If you have time to help review, these are open and waiting:

• #2167 — Add EarlyTranslate to EndpointSpec so the EPP receives an OpenAI-format body for Anthropic requests by @isztldav — discussed on the maintainer call; an upstream-chain alternative is being explored in parallel
• #2350 — Backfill input + cache token counts in OpenAI→Anthropic streaming message_delta by @ajac-zero — fixes the zero-input-token bug reported in #2352
• #2324 — Support for OpenAI Files API by @sivanantha321
• #2244 — Emit per-request rate-limit limits into dynamic metadata by @kanurag94
• #2169 — Add stream idle timeout with fallover to next model by @albe2669
• #2266 — Lua filters from EnvoyExtensionPolicy can now run after ext_proc on the response path by @yuehaii

🙏 Thanks to this week's contributors

@kanurag94, @walsm232, @pjdurden, @Aias00, @Killusions, @xiaolin593, @ajac-zero, @steffen-karlsson, @VadimEisenberg, @isztldav, @siddharth1036, @Hritik003, @0YHR0, @pbonuboinasnow, @sivanantha321, @albe2669, @yuehaii, @andrewrothstein, @nacx — and everyone who showed up to triage, review, and discuss.

See you next week!