PhotoCraft Is a Free Photoshop Rebuild an AI Agent Can Operate, and That Matters More Than Parity
An open-source, Photoshop-compatible editor shipped its first builds this weekend with a socket for AI assistants already built in. Here is what that changes for retouchers and designers, and what it still cannot do.
The makers of PhotoCraft estimate that only 20 to 35 percent of working professionals could switch to it today. They wrote that number into their own roadmap, next to a line admitting there is no generative fill at all. That is an unusual thing for a brand-new image editor to publish about itself in its first week. It is also the least interesting fact about the project.
The interesting fact is the door it leaves open. Every menu item in PhotoCraft, all 500-plus commands, can be run by an AI assistant in exactly the way you would run it by clicking. Not a chatbot that describes what to do. Not a panel that generates a picture and drops it on a new layer. The assistant opens your PSD, adds a real curves adjustment layer, applies a real smart sharpen, and saves a file that still has its layers.
Big editors have spent two years adding AI as a feature inside the app. PhotoCraft flips that around: the app is built to be a tool for the AI. My position is that this second approach matters more to anyone who does repetitive image work for a living, even though the editor itself is clearly unfinished.
What PhotoCraft actually is
PhotoCraft is a free image editor from the team behind ArtCraft, a desktop app for staging AI image and video shots. The repository was created on September 30, and the first downloadable builds for Mac, Windows and Linux arrived on October 2. It is written from scratch rather than copied from Adobe, which the team calls a "clean-room reimplementation."
The feature list reads like a Photoshop starter kit: layers, masks, 16 kinds of adjustment layer (curves, levels, vibrance and the rest), type and shape tools, layer styles such as drop shadows and strokes, and filters that stay editable after you apply them. It handles print color (CMYK and Lab) and high bit depths as well as ordinary RGB.
The headline claim is about files. According to the README, 134 of 135 real-world PSDs the team tested open and save back identical to the original, down to the last byte. Anything PhotoCraft does not understand yet is carried through untouched rather than thrown away. For anyone who receives PSDs from clients, that is the claim that decides whether a tool is usable at all, because an editor that silently drops a layer effect is worse than no editor.
The code is dual-licensed under MIT or Apache 2.0, both permissive, so you can use it on paid client work without asking anyone. The name and logo sit under a separate brand license, which only matters if you plan to fork and redistribute it.
The part launch coverage will skip: the agent socket
Here is the mechanism, in plain terms.
Most apps have two layers. There is what you see (menus, buttons, dialogs) and there is the code that does the work. In most editors those layers are tangled together, which is why automating Photoshop has always meant recording actions or writing scripts in Adobe's own languages.
PhotoCraft keeps a single list of every operation the program can perform. The menus call that list. So does a command-line version of the program, and so does a control channel for scripts. So does a small built-in server that speaks the Model Context Protocol, or MCP, a common standard that lets AI assistants such as Claude plug into a program and use its tools. You start that server with one line, photocraft-cli mcp, and an assistant connected to it can do anything the menus can.
Because there is only one list, nothing is second-class. The assistant is not using a reduced "API version" of the editor. It is pressing the same buttons you press, with the same results.
That has three practical consequences.
The output stays editable. When an assistant adds a curves adjustment through PhotoCraft, you get a curves adjustment layer, not a flattened image. Open the file later and you can drag the curve. Compare that with AI features that generate pixels onto a new layer: you can mask them, but you cannot tweak what they did.
A recipe you test by hand is the recipe that runs. Work out a look on one image, write down the commands, and the same commands run on the next fifty images. The README's own example applies a smart sharpen at amount 80 and a curves layer with four specific points, then exports a PNG, all from one line.
You own the workflow. It runs on your machine, against your files, with no subscription between you and your own process. If the project disappears tomorrow, the commands you wrote still describe exactly what you did.
Who this is for right now
Be realistic about the audience. PhotoCraft today is a good fit for three kinds of job.
The first is batch finishing. A photographer delivering 200 product shots that all need the same sharpening, the same tonal curve and a watermark layer. That is a loop over a folder.
The second is versioning. A designer who needs the same poster in six sizes and three color treatments, with every version kept as a layered file for the client. An assistant can open the master, make the change, save under a new name, and repeat, without anyone clicking through the same menus eighteen times.
The third is inspection. Ask an assistant to open a PSD and report what is in it: which layers are hidden, which fonts are used, which adjustment layers exist. That is useful before you send a file to print, and it does not change anything.
It is not a fit, yet, for painting, compositing that relies on content-aware tools, or anything where you want AI to invent pixels. Those tools are missing.
Put this into practice
Start small, and start with a copy of your files.
1. Install a build and test your worst file. Download v0.1.0 or the newer v0.1.1 pre-release from the releases page. Open the messiest PSD you have, the one with smart objects, layer effects and type. Save it under a new name. Open both in Photoshop and compare. If PhotoCraft mangles your real files, stop here; nothing else matters.
2. Run one finishing recipe from the command line. Try the README's example on a test file:
photocraft-cli run wave.psd --cmd filter.sharpen.smartSharpen --params '{"amount":80}' --cmd layer.newAdjustmentLayer.curves --params '{"points":[[0,0],[64,48],[192,212],[255,255]]}' --out wave-final.png
Swap in your own file and your own values. Once one image looks right, run the same line across a folder.
3. Connect an assistant. Start photocraft-cli mcp and add it as an MCP server in an assistant that supports them, such as Claude Desktop. Begin with read-only requests ("list the layers in this file and tell me which are hidden") before you ask it to change anything.
4. Write the recipe down. When an assistant gets a look right, ask it to give you the exact commands it ran. Save them. That text file is now a reusable action you own, which is the real payoff.
5. Check by eye. Before a batch goes to a client, open three or four outputs in the tool you trust and compare. A sharpening filter rebuilt from scratch will not match Photoshop's math exactly, and that difference shows up on skin and fine texture.
Where it breaks
This is early software, and its makers say so plainly. Take them at their word.
It is alpha. The README calls it "early alpha," and the roadmap says "real-file testing has only just begun," and that documents built like page layouts were laggy. Expect crashes. Keep originals.
One file in 135 did not survive. The round-trip claim is strong, not perfect. You do not know in advance whether your file is the one.
No generative tools at all. No generative fill, no expand, no neural filters, no remove tool. The roadmap lists about 20 Photoshop tools as missing entirely, including the pencil, the mixer brush, the patch tool and content-aware move. Stylus pressure is listed for Windows and the web, not for Mac or Linux.
"Parity" means the menus exist. The project's parity page reports 626 of 626 Photoshop menu items live, while the roadmap logged 532 of 625 on October 1, up from 224 the day before. As I read it, both track whether each command exists, not whether each one gives the same result as Adobe's version. The roadmap's own 20 to 35 percent estimate is the more honest number.
No Adobe ecosystem. No Photoshop action files or ExtendScript and UXP scripts carry over, no Adobe Fonts or cloud documents, and no classic .8BF plug-ins (the team plans its own sandboxed plug-in system instead). If your workflow depends on a plug-in, it stays in Photoshop.
An assistant can make a mess quickly. An AI that can run every command can also run the wrong one 200 times. Work on copies, start with read-only questions, and review before you overwrite anything.
What to watch
The question PhotoCraft raises is bigger than PhotoCraft. For years, automating creative software meant learning each vendor's scripting language, and AI arrived as one more feature behind a subscription. An editor designed from day one so that an assistant is a first-class user changes who can automate. A retoucher who has never written a script can describe a finishing pass in plain words and get back a recipe.
Whether PhotoCraft becomes a daily tool depends on months of work its makers have openly scoped. Whether the idea spreads depends on whether people try it on real jobs now.
So try it on one. Pick a dull, repetitive task you did last week, run it through PhotoCraft with an assistant on a copy of the files, and see whether the output survives a client's scrutiny. If it does, you have just bought back an afternoon. If it does not, you will know exactly which tool was missing, and that is worth filing on their GitHub.
Medium metadata
Title: PhotoCraft Is a Free Photoshop Rebuild an AI Agent Can Operate, and That Matters More Than Parity
Subtitle: An open-source, PSD-faithful editor shipped with an MCP server built in, so AI assistants use the same commands you do. What it can automate today, and the long list of what it still cannot do.
Tags: Photoshop, Photo Editing, AI Tools, Open Source, Design
Estimated read time: 8 minutes