Creative Work Just Moved Into the Chat Window, and It’s Not Leaving

I was scrolling LinkedIn when I stopped on a video unmistakably styled like the NBA Jam arcade game of my childhood, resplendent with pixelated, oversized athletes moving about the screen. Perhaps it was nostalgia that slowed my scroll, but it was the caption that gave me pause. Lenny Rachitsky, of the widely read Lenny's Newsletter, had posted it with this: "Now that is a launch video. I just told Claude Cowork to use Runway and create a launch video in the style of NBA Jam, promoting the Product Pass. Unreal."

I read it again: told Claude Cowork to use Runway. Not opened Runway. Not exported footage into Runway. Told Claude to use it.

Claude Cowork is Anthropic's agentic workspace for non-developers. Basically, you hand it a task in plain language, then it plans and executes the work itself. It coordinates whatever files, apps, and connected tools it needs along the way, without you touching each tool's native interface. In Rachitsky's case, that meant Cowork reasoning through the request, then reaching out to Runway — not as a human clicking through a video editor, but as a connected tool it could call directly — to actually generate and render the clip. The connective tissue making that possible is the Model Context Protocol (MCP), the open standard Anthropic released in November 2024 that lets AI systems discover and operate external tools through a common interface versus requiring a custom integration for every model and software pairing.

MCP, which was created at Anthropic by engineers David Soria Parra and Justin Spahr-Summers, donated the protocol oversight to the Agentic AI Foundation (AAIF) in December 2025. This summer's release, under AAIF's stewardship, was described as MCP's biggest update yet. This included moving to a fully stateless architecture, hardened authentication, and the graduation of MCP Apps. The latter is the piece that lets connected tools render real interfaces, not just text, directly inside a chat thread.

Marketing professionals across the spectrum — enterprise teams and boutique shops like mine — are taking note. This is the next big moment for marketing and design creatives since ChatGPT launched in 2022, which I covered for VentureBeat at the time. What's changed isn't that AI can edit video — that's been true for a while. It's that MCP connects generative AI chat to specialist tools behind the scenes, all driven by natural language intent. That means the chat window itself becomes the standard creative workbench. Goodbye tab-switching. Hello integrated, seamless creative.

Adobe Leading MCP Adoption

Adobe made one of the first big MCP splashes when it brought 50+ tools directly into Claude via its own "Adobe for creativity" connector. They announced a week after Adobe Summit wrapped in Las Vegas this past April.

For Adobe, this was a very measured decision, as they had spent the prior 18 months studying where creatives were actually using AI in their process. They identified how much of it started in a chatbot rather than inside Adobe's own ecosystem. That research surfaced in a June 2026 paper titled How AI Is Redistributing Creative Work. This was underpinned by a 1,433-person survey of professionals, as well as an eleven-country study of 3,300 people training to enter the field. What it found was high AI usage concentrated at the two ends of the design process. On the front end, 81.7% of pros rated brainstorming and ideation as having high AI enablement — the single highest score in the survey. On the back end, 57.6% of photo professionals reported substantial generative AI use for removing distracting background elements.

Adobe, a historically closed system, leading these headlines is not to be ignored. They are meeting users where they already are, and increasingly where the next generation starts their process by default — long before they ever open a native app.

Adobe's bet on chat-native distribution kept expanding from there. On August 6th, the company introduced the Adobe plugin in ChatGPT, consolidating more than 70 of its design and productivity tools into a single install. As Adobe put it in the announcement, the plugin is built for the "hundreds of millions of people or two million businesses using ChatGPT to brainstorm, research and solve problems." Now they can push those execution conversations further, straight into finished creative work from one platform — a shift the company showed off in a new announcement sizzle reel.

A Unified Prompting Interface, with Context and Creation Living in One Place

For decades, designers and marketers built expertise by mastering tools: Figma, Photoshop, After Effects, analytics platforms, as well as other research tools. Early AI workflows continued that same pattern. Teams had to learn the separate "prompt dialects" of Claude, ChatGPT, Runway, ElevenLabs, and whatever else sat in the stack — each with its own syntax for getting a good result. MCP is what breaks that pattern. Instead of learning a dozen dialects, you express intent in one chat interface, and the AI translates that intent into whatever the connected tool actually needs.

Some of what MCP also unlocks is rendering query answers beyond plain text. That's a specific extension called MCP Apps, which lets a connected tool hand back an actual interactive interface. It's available today in Claude on web, desktop, mobile, and in Claude Cowork. Where reviewing a generated video, adjusting a chart, or approving a design is critical in the process, MCP Apps is what makes the chat window function as a real workbench instead of a text log.

Instead of learning a dozen dialects, you express intent in one chat interface.

Practically speaking, a single Claude conversation becomes the marketing design gateway. For example, when I make videos and want to add a custom AI voice, I have to work the prompting in Runway, upload the ElevenLabs voice that I created there, and then combine them back in Runway. Using MCP, I can call Runway and ElevenLabs in one command, with the rendered result showcased back in that same thread. It preserves a project's context, including the brief, the brand voice, and the last three rounds of feedback attached to the work vs. scattered across five open browser tabs. This streamlining cuts the tab-surfing that eats the middle of every creative process.

MCP can also connect chats to prior campaigns and competitive research to inform a new direction. Imagine your chat querying a shared drive for past creative, checking a Slack channel where a client's feedback history lives, scraping competitor sites via Tavily or Firecrawl for a differentiated angle, and then confirming the results against Semrush. Now, that doesn't require an enterprise data stack. A small shop running this off a shared Google Drive and a couple of Slack channels gets a contextual throughline that can rival what a large in-house team gets off a formal database.

What It Costs, and What It's Worth

With data moving between systems, security and cost are the real questions before adoption. On security, MCP's biggest recent update hardened authentication against OAuth mix-up attacks. The extension's documentation is direct about the containment model: "MCP Apps run in a sandboxed iframe controlled by the host. They can't access the parent page, steal cookies, or escape their container." While it covers the interface layer, it doesn't cover what a connected tool does with the data you hand it, which is the question worth asking before connecting anything to client work.

Also, the ease of use still varies a lot by connector. Some install and authenticate in under a minute; others need real configuration before they're usable. That gap should close as more tools adopt the standard.

Cost is the second major consideration, which for now is pretty straightforward: every connected tool adds tokens on top of whatever the tool itself charges. For instance, Claude's usage is added to Runway's render cost. Whether that math works depends on what it's replacing.

Enterprises and small shops are both winning from this evolution, just from opposite directions. A 50-person in-house team and a five-person shop are chasing the same brief-to-delivery speed. However, they've been paying for it differently. The enterprise pays in bloat: a stack of licenses, seats, and admin overhead that scales with headcount, not output. The five-person shop pays in absence: no ops layer at all to manage that many separate logins. MCP collapses the same tool-switching overhead from both directions at once, letting the enterprise trim its sprawl down to one interface, and lets the small shop build capability it never had the staff to support.