Runway MCP, Claude Code v2.1.152, OpenAI Private MCP tunnels

May 27, 2026 is marked by a convergence around the MCP protocol: Runway launches an MCP server to connect its video and image models to Claude, ChatGPT, and Cursor, while OpenAI secures private MCP servers in the enterprise. At the same time, Claude Code v2.1.152 arrives with automatic code fixes, Anthropic opens an office in Seoul, and Qwen3.5 sets a speed record at 580 tokens/s.

Runway MCP — Video and image generation from Claude, ChatGPT, Cursor

May 27 — Runway launches an MCP (Model Context Protocol) server that makes it possible to integrate its video and image generation capabilities directly into developers’ and creatives’ work environments. Setup takes just a few seconds: in Claude Desktop or claude.ai, go to Customize → Connectors, add the https://mcp.runwayml.com/mcp connector, then sign in with a Runway account.

The distinctive feature of this integration is the diversity of models accessible through a single entry point. Runway does not expose only its own models, but aggregates several third-party engines within the same MCP server:

Model	Type
Gen-4.5	Video
Gen-4 Turbo	Video
Seedance 2.0	Video
Kling 3.0	Video
Veo 3.1	Video
GPT Image 2	Image
Nano Banana Pro	Image
Gen-4 Image	Image

Compatible agents: Claude.ai, Claude Desktop, ChatGPT, Cursor, Replit, and any agent supporting the MCP standard.

Among the illustrated use cases: generating product images from a URL, creating dialogue-driven video ads, adding visuals for a marketing site. Billing is based on the usual Runway credits, and the feature is available immediately for all teams.

“Introducing Runway MCP. Now you can connect Runway directly into Claude, ChatGPT, Cursor, Replit and more. Generate polished images and videos with state-of-the-art models, like Gen-4.5, Seedance 2.0, GPT Images 2.0, Kling and more. Right from where you’re already working.”

Introducing Runway MCP. Connect Runway directly into Claude, ChatGPT, Cursor, Replit, and more. Generate polished images and videos with state-of-the-art models like Gen-4.5, Seedance 2.0, GPT Images 2.0, Kling, and more. Directly from where you already work. — @runwayml on X

🔗 Runway MCP product page

Claude Code v2.1.152 — /code-review —fix, MessageDisplay hook, Auto Mode

May 27 — Claude Code version 2.1.152 ships a set of features focused on automated code review, hook extensibility, and plugin management.

New feature	Description
`/code-review --fix`	Automatically applies review suggestions (bugs, simplifications, efficiency) in the working directory. `/simplify` is now an alias.
Hook `MessageDisplay`	New event type allowing extensions to transform or hide the text displayed on screen.
Enhanced `SessionStart`	Return `reloadSkills: true` to rescan skills directories without restarting, set the session title via `hookSpecificOutput.sessionTitle`.
`/reload-skills`	Reloads skills without restarting the session.
Auto Mode	No longer asks for explicit opt-in confirmation.
`--fallback-model`	Automatically switches to the fallback model if the primary model cannot be found, instead of failing on every request.

Among the notable fixes: terminal stability during long sessions, the “still thinking” spinner displayed in error, MCP servers with the same command but different environment variables deduplicated incorrectly, and sessions stuck after a model change.

🔗 Claude Code CHANGELOG

Anthropic opens an office in Seoul — KiYoung Choi named Representative Director Korea

May 26 — Anthropic announces the appointment of KiYoung Choi as Representative Director for Korea, ahead of the opening of an office in Seoul. Korea is described as one of the most active markets for Claude.ai: according to Anthropic’s economic index, Koreans use Claude at 3.5 times the expected rate for the population size.

KiYoung Choi joins Anthropic from Snowflake, where he was General Manager for Korea. He brings more than 30 years of experience leading technology companies in Korea and Asia-Pacific, with previous positions at Google Cloud, Adobe, Autodesk, and Microsoft. Korean companies already among its customers include Law&Company (AI legal assistant) and SK Telecom (customer service model built on Claude). Anthropic’s executive team will travel to Seoul in the coming weeks for the official inauguration.

🔗 Anthropic announcement

OpenAI Developer Updates — Private MCP tunnels, Workload Identity Federation, Admin API

May 27 — @OpenAIDevs publishes a three-part thread announcing new features for enterprise developers on the OpenAI platform.

Private MCP servers: teams can now keep their MCP servers inside their private network while making them accessible to ChatGPT, Codex, and the Responses API via secure tunnels.

Workload Identity Federation: a cloud-based identity for the OpenAI API platform. Teams manage access through IAM (Identity and Access Management) workflows, reducing the need to distribute permanent API keys across services.

Expanded Admin API: the administration API is expanded to allow companies to manage their OpenAI projects programmatically.

🔗 OpenAI Devs thread

Google AI — Gemini Omni prompting guide (video generation)

May 26 — Google AI publishes a guide to five techniques for getting the most out of Gemini Omni Flash in video generation:

Real-world knowledge — Gemini understands history, science, and culture; cultural and scientific references are enough without describing everything.
Text rendering — Integration of text into videos with precise typography, positioning, and animation options.
Camera direction — Response to cinematic instructions (dolly zoom, push in, camera styles such as vintage film or grainy webcam).
Iterative editing — Targeted modifications (changing a background, adjusting a subtitle) without rewriting the entire prompt.
Changing the action on the fly — Change a character’s pace or emotions without breaking continuity.

Gemini Omni Flash is available in the Gemini app, Flow by Google, Google Flow Music, and YouTube Shorts and Create.

🔗 @GoogleAI post

Genspark AI Slides 5.0 — Slide Skills and 5x performance

May 27 — Genspark announces version 5.0 of its AI Slides tool, presented as the biggest upgrade since launch. The central new concept is Slide Skills: reusable blocks that encapsulate expertise logic and visual style, applicable with one click.

New feature	Description
100+ built-in Expert Skills	Consulting, sales, teaching expertise — applicable with one click
Build your own Skill	Upload a deck or PDF + description of analysis steps
5x performance	New architecture and more powerful models
Integrated data analysis	Reads Excel and Google Sheets files, transforms them into slides

🔗 Genspark AI Slides 5.0 announcement

FastVideo Dreamverse — 30 seconds 1080p in 7s on Blackwell GPU, open source

May 27 — The Hao AI Lab (UCSD) open-sources FastVideo Dreamverse, a real-time video generation application reaching 30 seconds of 1080p video in 7 seconds on a single NVIDIA B200 (Blackwell) GPU. By comparison, it previously took about 25 seconds to generate 5 seconds of video on 8 Blackwell GPUs — an efficiency gain of about 48x.

The optimizations rely on FA4 (Flash Attention 4 by @tri_dao), the LTX-2 base model, and contributions from NVIDIA AI, MBZUAI, and the Mooncake team. NVIDIA AI amplified the announcement by highlighting the shift from 25s for 5s of video on 8 GPUs to 4.2s on a single Blackwell GPU.

🔗 @haoailab tweet

Qwen3.5 — Record 580 tokens/s on TokenSpeed (NVIDIA/PyTorch)

May 27 — The Qwen team announces that Qwen3.5-397B-A17B achieves a record 580 tokens per second for agentic workloads on the TokenSpeed inference engine, developed in collaboration with NVIDIA and the PyTorch Foundation.

Engine	Model	Performance
TokenSpeed	Qwen3.5-397B-A17B	580 tps (record)

This result is made possible by FA4 (Flash Attention 4) optimization from @tri_dao, with contributions from Lightseek.org, NVIDIA, and the Mooncake team. The entire work is published as open source.

🔗 @Alibaba_Qwen announcement

Briefs

Claude Marketplace — Anthropic adds 5 new partners: Augment Code, Bolt, CodeRabbit, Hebbia, and Legora. Enterprise customers can allocate their Anthropic commitment to these Claude-powered products. 🔗 tweet @claudeai
Gemini Spark + Daily Brief — Google Gemini hosts a community Discord event on May 27 to present these two agentic features announced at Google I/O, with live demos and Q&A with the team. 🔗 tweet @GeminiApp
ElevenLabs — Stan Lee voice — The official voice of the Marvel creator is available on the ElevenLabs Iconic Marketplace and in ElevenReader. 🔗 tweet @ElevenLabs
Grok in Kilo Code — xAI announces Grok integration in Kilo Code, the open-source agentic coding platform. Accessible to SuperGrok and X Premium+ subscribers. 🔗 x.ai/news
Tax AI with Codex (Thrive Holdings / Crete) — OpenAI publishes a case study of Tax AI, an automated tax preparation system for the Crete network (30+ firms). Results: 7,000 pilot returns, accuracy up to 97%, and about one-third time savings for accountants. 🔗 openai.com
GPT-5.5 in Codex for Databricks — OpenAI announces that GPT-5.5 in Codex helps Databricks analyze complex customer documents with greater reliability. 🔗 tweet @OpenAIDevs
Perplexity open-sources its Unigram tokenizer — 5 to 6 times lower CPU latency compared with HuggingFace tokenizers, thanks to the elimination of string reconstruction and hash table lookups. Published in repository plx-garden. 🔗 tweet @perplexity_ai

What it means

MCP as a multimodal orchestration protocol. The launch of Runway MCP marks a step in MCP adoption beyond text and code. By exposing eight video and image models through a single standard entry point, Runway turns a creative tool into an agentic infrastructure building block. The fact that OpenAI simultaneously announces private MCP servers in the enterprise confirms that MCP is becoming the reference protocol for connecting tools to AI agents — both for consumer use and for secure enterprise deployments.

Developer tooling: automation and extensibility. Claude Code v2.1.152 crosses a threshold with /code-review --fix: code review moves from commentary to direct action in the working directory. The MessageDisplay hook opens a new layer of extensibility by allowing response display to be transformed on the fly. These two additions fit into a trajectory where the development assistant becomes an agent capable of acting, not just advising.

AI geopolitics: Anthropic in Korea. Opening an office in Seoul with an experienced representative director (ex-Snowflake, Google Cloud, Adobe) reflects a local anchoring strategy in markets with strong adoption. The figure of 3.5 times the expected usage rate places Korea among Anthropic’s strategic markets in Asia-Pacific, alongside the partnerships already established with SK Telecom and Law&Company.

Open-source inference performance. Qwen3.5’s 580 tokens/s record and the release of FastVideo Dreamverse (30s of 1080p video in 7s on a single Blackwell GPU) illustrate the convergence between open academic research and NVIDIA’s latest-generation hardware optimizations. The NVIDIA/PyTorch/Mooncake collaboration on TokenSpeed and the FA4 optimizations on LTX-Video show that the most significant performance gains are now emerging from collaborative ecosystems rather than isolated proprietary developments.