AI Engineering Digest

Hybrid Apigee speaks MCP.

Tuesday, September 15, 2026

Issue 17 · sourced briefing

Sources: Apigee hybrid RN · llama.cpp · mlx-vlm · Databricks · GC RN

Breaking Apigee hybrid v1.17.0 adds MCP support — hybrid-managed APIs become agent tools via managed MCP endpoints (optional, off by default)

Lead Story

Hybrid Apigee speaks MCP — managed endpoints for on-prem API estates

Apigee hybrid v1.17.0 (September 14 release notes) adds Model Context Protocol support so agentic apps can call hybrid-managed APIs as tools through managed MCP endpoints — without running your own MCP servers.

Google Cloud aggregate and Apigee hybrid release notes under September 14, 2026 ship hybrid v1.17.0 with MCP support. The RN says Apigee hybrid routes, authorizes, and secures MCP tool calls the same way it manages other APIs. Support is optional and not enabled by default — you must explicitly enable it in hybrid configuration.

This is distinct from Issue 16’s cloud API Gateway remote MCP Preview (REST→tools on Gateway). Hybrid is the on-prem / customer-cluster runtime. Cloud Apigee MCP overview pages still carry banners that MCP in Apigee is not for hybrid organizations — treat the dated hybrid RN as the primary for hybrid MCP; do not conflate the two. Hybrid enable/quickstart paths named in the RN resolved poorly this research pass — cite the dated RN until those docs stabilize. vendor

Why it matters: hybrid API estates get a control-plane path to agent tool surfaces with familiar Apigee routing and auth, as a follow-on to yesterday’s cloud Gateway MCP story without duplicating it.

Sources: Apigee hybrid release notes · Google Cloud release notes

Today’s Stories

Local · Apple Silicon

Metal FA unlocks MiniCPM3 in llama.cpp b10978

Nightly adds flash-attn kernels for HSK=96 / HSV=64 so MiniCPM3 no longer aborts on -fa auto. Fixes the Metal missing-kernel crash. Contributor benches are M5 Max only — do not treat as Studio / M3 Ultra tok/s. contributor

Source: b10978 · #28599

MLX · multimodal

mlx-vlm 0.7.1 GA for Apple Silicon VLMs

Patch multimodal release after 0.7.0: Hy4, GLM-5 Next, Spark-X2.5, SAM 3.1, Mage-VL video, plus a Metal-timeout SFT path fix. This is mlx-vlm only — mlx and mlx-lm unchanged.

Source: v0.7.1 · PyPI

Enterprise · security

Claude Cyber Verification Program Preview on GCP

Anthropic’s CVP is in Preview on Gemini Enterprise Agent Platform (Sep 14): verified orgs can lift default dual-use restrictions for legitimate defensive cyber work. Enrollment + Anthropic approval required — not a general Claude unlock. vendor

Source: CVP docs

Enterprise · Databricks

Genie charts now render in Slack replies

Public Preview Sep 14: when a Genie response includes a chart, the visualization appears as an image in the Slack reply. Prior Genie Slack installs need a reinstall; private visibility still hides images.

Source: Databricks RN

Local · Ollama RC

Ollama v0.34.1-rc2 — MLX runner memory RC

Prerelease only (stable still v0.34.0): mlxrunner prefix-cache eviction, free-memory waits before loading the next MLX model, scoped array lifetimes. Label as RC — not GA.

Source: v0.34.1-rc2

Signal Tables

Model releases (this week)

Model / package	Org	Notes
mlx-vlm 0.7.1 Sep 14 GA	Blaizzy	Open Multimodal MLX — not mlx / mlx-lm
MiniCPM3 Metal FA b10978 · Sep 15	llama.cpp / openbmb	Open HSK=96 / HSV=64 kernels; -fa auto unblocked
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

Security / runtimes

Item	Beat	Status
Claude CVP Sep 14 Preview	Defensive cyber on Agent Platform	Watch Pre-GA · enrollment
Apigee hybrid MCP v1.17.0 · Sep 14	Hybrid APIs → agent tools	Optional Off by default
llama.cpp b10978 Sep 15 ~6:59 AM ET	Metal FA MiniCPM3	Nightly
vLLM 0.29.0 / SGLang 0.5.19 Standing	Serving runtimes	GA

Industry Shorts

Critical · MiniCPM3 on Metal

If MiniCPM3-4B crashes on Apple Silicon with -fa auto (missing dk96/dv64 kernel), pull llama.cpp ≥b10978. Contributor speedups are M5 Max — not a Studio claim.

Watch · Hybrid vs cloud MCP

Apigee hybrid MCP is a separate ship from Issue 16’s API Gateway / cloud Apigee MCP. Cloud docs still say not for hybrid orgs — cite the Sep 14 hybrid RN; enable docs were shaky this pass.

Opportunity · mlx-vlm 0.7.1

Mac multimodal stacks can upgrade to 0.7.1 for Hy4 / GLM-5 Next / Spark-X2.5 / video paths while leaving mlx and mlx-lm on standing tags.

Try This

Run MiniCPM3-4B with -fa auto on llama.cpp ≥b10978

Pull macOS arm64 nightly b10978+ (or build after #28599). Use openbmb MiniCPM3-4B GGUF Q4_K_M. Confirm -fa auto no longer aborts; optional benches are contributor M5 Max only.

local metal b10978

pip install 'mlx-vlm==0.7.1'

Upgrade from 0.7.0 for the multimodal patch release (requires standing mlx≥0.32.2). Confirm with python -c "import mlx_vlm; print(mlx_vlm.__version__)" → 0.7.1. mlx / mlx-lm stay put.

mlx-vlm apple silicon 0.7.1

Issue 17 · Tuesday, September 15, 2026 · Not investment advice · Sources cited inline · Other news in AI: none this window · Frontier / OSS tools / models / field: quiet