AI Engineering Digest

API Gateway speaks MCP.

Monday, September 14, 2026

Issue 16 · sourced briefing

Sources: Google Cloud RN · llama.cpp · AWS · Databricks

Breaking Google Cloud API Gateway acts as a remote MCP server (Public Preview) — REST estates become agent tools without backend rewrites

Lead Story

API Gateway speaks MCP — turn REST into agent tools without rewriting backends

Google Cloud puts API Gateway in Public Preview as a remote Model Context Protocol server: annotate OpenAPI 3.x, keep your backends, and let agents discover and call tools through the gateway.

Aggregate release notes dated September 11 ship Enable Model Context Protocol (MCP) on API Gateway. The gateway receives MCP JSON-RPC over HTTP POST and translates to REST toward existing services. Lifecycle methods covered in Preview: initialize, notifications/initialized, tools/list, and tools/call. Resources and prompts methods return JSON-RPC method-not-found.

Config is OpenAPI 3.x only (2.0 unsupported): global x-google-api-management.mcp: true and/or per-operation x-google-mcp-tool. Mutually exclusive with Model Routing in the same API config. Auth nuance: lifecycle methods are unauthenticated; tools/call reuses operation security; docs strongly recommend JWT for tools/list. MCP-enabled gateways can publish to API Hub and show up in Agent Registry when integrated.

Why it matters for AI engineers: this is a control-plane path to expose an existing REST estate as agent tools with gateway auth and catalog discovery — distinct from sandbox/VPC-SC Agent Gateway work already covered in prior issues. Status is Public Preview (Pre-GA terms). vendor

Sources: Google Cloud release notes · MCP overview · Configure MCP

By the Numbers

Enterprise ships in this window (MCP Gateway, Agent Substrate, DevOps Slack, Unity Gateway Flash)

b10955

llama.cpp nightly fixing macOS arm64 heap corruption after b10909 (~6:42 AM ET Sep 14)

New frontier flagship API ids · MLX GAs · OSS tool/model GAs this window

Local regression window

Reporter: last good b10909; broken from source around b10919+. Root cause: ggml-cpu PCH vs 256/64 CACHE_LINE_SIZE mismatch. Fix merged in #28882. Formula/undersize claims from PR text — not an independent Mac Studio bench. contributor

Today’s Stories

Local · Apple Silicon

Mac heap fixed in llama.cpp b10955

ggml-cpu PCH force-included headers that set CACHE_LINE_SIZE to 256 on Apple Silicon while C work buffers used 64 — undersized rope buffers, heap overflow, decode crashes. Disable PCH + deterministic CACHE_LINE_SIZE. Last good was Issue 15’s b10909. contributor

Source: b10955 · #28858

Enterprise · GKE

Agent Substrate on GKE — density for idle agents

Sep 11 RN: eval / non-production now; production on allowlist limited GA. Suspends idle agents, snapshots RAM + local files, restores with sub-second latency claims — denser than Agent Sandbox via bypassing standard K8s control-plane bottlenecks. Concept docs 404 this pass; cite RN. vendor

Source: Google Cloud RN

Enterprise · AWS

DevOps Agent: bidirectional Slack investigations

What’s New Sep 11 (~11 AM ET): manage the full investigation lifecycle in a private Slack channel — @mention to start/steer; findings stay in-thread. Docs flag LLM findings may be inaccurate. Distinct from Gemini Enterprise Slack mentions (Issue 14).

Source: AWS What’s New

Enterprise · Databricks

DeepSeek V4.1 Flash on Unity Gateway

Sep 11 product + Unity Gateway notes: V4.1 Flash available with text and image inputs via Foundation Model APIs (databricks-deepseek-v4-1-flash). Enterprise hook is governed gateway availability — model lab story was Issue 14.

Source: Databricks RN

Signal Tables

Model releases (standing)

Model	Org	Notes
DeepSeek-V4.1-Flash Issue 14 · Unity Gateway Sep 11	DeepSeek / Databricks	Cloud On Unity Gateway + Foundation Model APIs
gpt-6-astra Standing	OpenAI	Cloud No new flagship text id this window
claude-fable-5-1 Standing	Anthropic	Cloud Unchanged (Issue 9)
grok-4.6 Standing	xAI	Cloud No grok-4.7 on docs

Runtimes / security

Item	Beat	Status
llama.cpp b10955 Sep 14 ~6:42 AM ET	macOS arm64 heap / ggml-cpu PCH	Fixed
vLLM 0.29.0 Standing	Serving runtime	GA
SGLang 0.5.19 Standing	Serving runtime	GA
API Gateway MCP Sep 11 Preview	REST → agent tools	Watch Pre-GA

Industry Shorts

Critical · Mac nightlies

If you built llama.cpp after b10909 on Apple Silicon and saw decode SIGTRAP / EXC_BAD_ACCESS in rope — upgrade to b10955 or rebuild past #28882. Workaround was CMAKE_DISABLE_PRECOMPILE_HEADERS=ON.

Watch · MCP on Gateway

Public Preview: tools/list auth defaults are weak — docs push JWT. resources/* and prompts/* not supported yet. Don’t treat as production MCP platform GA.

Opportunity · REST estates

Teams with mature OpenAPI 3.x inventories can trial agent tool surfaces without backend rewrites — pair with Agent Registry / API Hub discovery when integrated.

Try This

Upgrade llama.cpp to b10955 on Apple Silicon

Pull llama-b10955-bin-macos-arm64.tar.gz or rebuild past #28882 if post-b10909 builds crash on decode. Confirm a short decode on a model you already use — reporter CI is not an M3 Ultra bench table. Optionally watch open follow-on #28892 (full PCH removal).

local macos arm64 b10955

Sketch an MCP-annotated OpenAPI 3.x on API Gateway

If you already have Gateway + OpenAPI 3.x, read the MCP overview/configure docs and trial tools/list / tools/call against a non-prod API. Flag Pre-GA terms; harden tools/list with JWT.

enterprise mcp preview

Issue 16 · Monday, September 14, 2026 · Not investment advice · Sources cited inline · Other news in AI: none this window · Frontier / MLX / OSS tools / models / field: quiet