Adobe will let ChatGPT, Claude and Gemini work directly inside Photoshop and Premiere

Adobe is opening four of its biggest creative apps to AI agents it doesn't make. Assistants from OpenAI, Google and Anthropic will be able to work inside Photoshop, Illustrator, Premiere and After Effects and change a user's actual project files.

The company announced the plan Tuesday at its Adobe MAX conference in Miami, along with a wide set of AI features across Creative Cloud and its Firefly platform. The third-party connection uses the Model Context Protocol, or MCP, an emerging standard for linking AI models to outside software. It is not available yet. Adobe says a private beta will open in early 2027 and is taking sign-ups from interested users.

That date matters more than anything else in the announcement. The most strategically interesting thing Adobe showed is still more than a month from even a limited test, while most of what users can try this week is a long list of smaller editing upgrades.

How outside agents will work in Creative Cloud

Adobe's MCP integration runs locally on the user's computer instead of through a cloud upload. An agent such as ChatGPT or Claude would work on the native file, so Photoshop layers, Illustrator vectors and Premiere bins and sequences stay intact and editable. A person can step in at any point and finish the work by hand.

Agents connected to other MCP-capable services can fold those into the same job. In Adobe's examples, a user could have an agent pull reference material from Google Docs, build a design in Adobe's apps and then post the result to colleagues in Slack. Another example has ChatGPT working with Photoshop to adapt a whole marketing campaign around a new piece of artwork attached to a prompt, then dropping the files into a Google Drive folder.

Adobe already offers ChatGPT and Claude integrations that reach some of its product functions. The local MCP is meant to open up far more of the apps. Deepa Subramaniam, an Adobe vice president, told TechCrunch the company hasn't settled how much to expose and will watch how people use the first set of capabilities.

"We are still figuring out where and when and how to draw the line," she said. She added that Adobe will add more over time and said, "We don't have any need or desire to sort of limit what the MCP endpoint can expose."

The logic behind the move looks sound. Plenty of designers and marketers already spend their day in a general-purpose chatbot, and Adobe can't force them to switch to its own assistant. By keeping the work in Photoshop and Premiere files, Adobe makes those formats the place where finished work ends up, whichever AI did the clicking. Protecting that position probably matters more to Adobe than winning a chatbot race against much larger AI labs.

The risk is who gains the most from the arrangement. Last month, at an event in India, Figma co-founder Dylan Field said MCP is not a perfect standard and benefits frontier labs more. If ChatGPT or Claude becomes the screen where creative work is directed, Adobe's apps start to look like a back-end service. Subramaniam's open-ended answer suggests Adobe hasn't decided how far it will go toward that model.

Anil Chakravarthy, who becomes Adobe's CEO on Dec. 1, demonstrated a third-party agent onstage cutting a long-form podcast into several social media clips. That kind of tedious repurposing is exactly where agents make sense. Nobody gets into video editing to slice one episode into a dozen vertical clips.

Adobe's own agent spreads across the apps

Adobe isn't handing the agent role to outsiders. Its Creative Agent, introduced earlier this year, now powers AI assistants in beta in Illustrator, Premiere, InDesign, After Effects and Frame.io, plus the Photoshop beta app. The assistants take plain-language requests and carry out multistep jobs. In Premiere, for example, an editor can ask one to apply an effect to a whole sequence, follow a masked subject across frames, or match color and exposure across many clips.

Premiere's mobile app is also getting agent tools soon. Quick Cuts will build a rough first edit from a prompt and the footage a user adds, and an AI Assistant will find and run editing tools from typed instructions.

Adobe creativity chief David Wadhwani said in April that the creative agent would eventually appear "in virtually everything we develop." Tuesday's lineup makes clear he meant it.

The Firefly AI Assistant beta adds Plan Mode, which shows the agent's proposed steps so a user can review, edit and approve them before anything runs. A Model Selector lets users pick an AI model or leave the choice on automatic. Users can also store preferred formats, colors, brand guidelines, characters and products and apply them across projects, and the assistant now connects with Firefly Boards.

Plan Mode may be the most practical addition of the day. An agent that changes a client file without showing its work is a liability, and a preview-and-approve step is the minimum professionals should expect.

Adobe also previewed "always-on" agents that handle recurring work in the background, such as sorting incoming assets, running quality checks or turning new images into graphics sized for different social platforms. A user would check in, approve the agent's plan and choose the best results. These will arrive first in Firefly AI Assistant, but Adobe gave no launch date.

Firefly edges toward Figma's turf

Firefly is growing well beyond image generation. Firefly Builder, in private beta, prototypes websites and apps from natural-language descriptions. Firefly Graph lets creators build reusable, repeatable workflows, and Firefly Design combines illustration, image editing, animation, 3D and print on one canvas. A new Firefly Video tool organizes scenes, assembles clips, adds captions and generates transitions and B-roll. Quick Actions, preset workflows for jobs such as making thumbnails or product ads, are now generally available.

The Verge noted that Builder resembles Figma Make and that Graph looks like Figma's Weave platform. Builder in particular puts Adobe in direct competition with Figma for early-stage product design, which helps explain why Field's criticism of MCP lands as more than a technical aside.

Editing upgrades users can try now

Much of the immediately useful news is less flashy. Photoshop gains Relight, which places movable light sources in a scene with adjustable hue and brightness, and Generative Restyle, which turns prompt-based style changes into sliders. A Quick Compare slider splits the canvas for before-and-after checks, and the updated Remove Tool now also cleans up related reflections, shadows and footprints. Topaz Labs, which Adobe acquired, is adding two image models: Wonder 4, which combines upscaling, denoising, deblurring and sharpening, and Image HDR, which converts standard images to HDR.

Premiere adds Enhance Audio, Separate Crosstalk and Separate Sound Source, which split dialogue, music, ambience and effects onto their own tracks. Generative Remove, now in beta, erases objects such as boom mics or logos from footage after a user masks and tracks them. The Color Mode grading tool, which entered beta in April, is now generally available, as is the Generative Media Tool. That tool creates video, sound effects, soundscapes and music on the timeline using Firefly or partner models from Google, Kling, Runway and Luma.

Premiere mobile users can now use hundreds of thousands of licensed tracks from Universal Music Group and Warner Music Group, for noncommercial use only.

Lightroom adds Glow for a soft, cinematic light effect, brings one-tap blemish removal to desktop and expands camera tethering, including connecting cameras directly to an iPad. Open Eyes, in public beta, replaces closed eyes with open ones by drawing on a reference photo of the same person from the user's library.

Illustrator gets new brushes, gradients, patterns and Blend controls with a live preview. InDesign adds Co-editing in beta, a Flex Layout tool that adapts content to canvas size, and Check Document and Rewrite for catching editorial and accessibility problems during work.

Open Eyes and Generative Remove are the kind of fixes photographers and editors have done by hand for decades, and Adobe is wise to keep them in beta while it tests how well they hold up. The bigger story, though, remains the MCP connection. Until the 2027 beta shows how much of Photoshop and Premiere Adobe actually exposes, the promise of outside agents doing real work in those apps is still a demo.