Adobe opens Photoshop and Premiere to third-party AI agents via MCP

Imagine asking Claude or ChatGPT to put together banners from a brief in Google Docs. Instead of sending you a flat image, the agent opens Photoshop, works with layers and sends the finished result to Slack. That is the direction Adobe is pushing Creative Cloud in. The company laid out its plans at the Max conference in Miami.

Third-party agents inside Photoshop

The main news is local integrations via Model Context Protocol (MCP), an open standard that lets AI agents connect to external tools. Adobe showed such integrations for Photoshop, Illustrator, Premiere and After Effects. An agent from OpenAI, Google, Anthropic or another provider will be able to work on your computer and make changes directly in native project files.

An important detail is that the result is not "baked" into a raster image. Layers, vectors, bins and sequences remain editable, and a designer can take over at any moment and refine things by hand. Most generative tools hand over a finished picture that is often easier to redo from scratch than to fix. Adobe promises the opposite here.

The second part of the scenario is chaining. If an agent is also connected to other MCP tools, it can pull references from Google Docs, assemble a layout in Adobe apps and send the result to colleagues or a client in Slack. In this setup, Creative Cloud becomes one link in someone else's agent pipeline rather than a closed sandbox.

The timeline is modest. The closed beta of local MCP integrations will launch in early 2027, and for now it is only a preview.

Agents that work while you are busy with something else

The second announcement is "always-on" agents. They are assigned repetitive work: sorting incoming assets, constantly checking quality, preparing content. The user reviews the agent's plan and approves the final result.

Adobe's example: an agent automatically turns all incoming images into promo graphics for different social networks. You look in once a day, pick the best options and publish. The company did not name a preview date, but it will arrive first in Firefly AI Assistant.

What's new in Firefly AI Assistant

The beta version of the assistant, built on Adobe's creative agent, received several updates:

— Plan Mode: you can examine, edit and approve the agent's plan before work begins. — Model Selector: you can pick a model for a project or leave automatic selection based on the desired outcome. — Integration with Firefly Boards. — Saved preferences: formats, colors, brand guides, characters and products are applied to all projects automatically.

Since the beginning of the year, the creative agent has been built into Premiere, Photoshop, Illustrator, InDesign, After Effects and Frame.io. Now it goes deeper into video editing. In Premiere, it can track masked objects in the frame, apply effects, and adjust color and exposure across entire sequences of clips at once. The agent will also come to mobile Premiere, so you will be able to edit by conversation while on the go.

What this means in practice

If you are building a product with agents, there are two patterns worth borrowing from Adobe. The first is editable output: the agent works in the native format, and a human can always pick up the task. The second is Plan Mode: autonomy without the ability to check the plan quickly runs into user distrust.

The strategic move is also clear. People already live in their favorite agents and do not want to retrain on a built-in chat. By opening MCP, Adobe is trying to remain the system where creative files live, even if the interface on the outside is someone else's. For studios and product teams, this is a chance to connect an already familiar agent to the design pipeline without new assistant subscriptions.

The limitations are obvious too. Right now, none of the MCP part of the announcement can be tried, and background agents do not even have a preview date. The material does not answer questions that will arise immediately: how to restrict the permissions of a third-party agent writing to working files, who pays for generations, and how to verify the result after long overnight work. These are our own thoughts, not Adobe's statements.

In April, David Wadhwani, head of the creative division, said the creative agent would appear in "practically everything we develop." Max showed that this is not mere rhetoric. What remains is to wait for the beta in 2027 and see how open the doors really are to other companies' bots.