
Unofficial project · Not affiliated with or endorsed by Adobe Inc.
Most AI products start by putting a chat box in front of a new system.
The more interesting problem, at least for me, is the opposite: people already have a system they trust — Photoshop, in this case — and the agent still cannot reliably drive it. Creative Cloud is not going anywhere. The gap is the last metre: natural language in, pixels out, files never leaving the machine.
In February I published a small MCP server: fifty-odd tools, npx, and that idea. I thought a handful of people who already lived in both Cursor and Photoshop might try it.
They did. Then a lot of other people did too.
Top 200 of the MCP world
MCP Toplist tracks 140,843 servers. Photoshop MCP is #139 — inside the top 200. GitHub stars went from almost nothing in April to about 567 now.
The same listing shows up on the other boards people actually use:
- PulseMCP estimates around 30,000 visitors and ranks it #941 globally — #242 this week.
- It is on the Official MCP Registry as
io.github.alisaitteke/photoshop-mcp(latest 1.7.21). - Glama carries the same package.
On npm it is pulling about 10,400 downloads a week.
Those are directory scores. Useful for discovery. The number I care about more is who actually opens Photoshop with it — because leaderboards are easy to game, and sessions are not.
About 1,200 people a day
Over the last two weeks, anonymous usage from the server itself shows about 1,200 people a day starting a session — connecting an AI client to Photoshop through this MCP.
They are not all in one city. The busiest countries in that window: China, the United States, Japan, then South Korea, Brazil, and Germany. That was not a launch-market I planned. It is what you get when the install is npx and the product is a tool people already pay Adobe for.
Somewhere, it’s always a workday
The six countries that open Photoshop MCP most, and the hours their offices are open — redrawn on your clock.
Between them, these offices cover every hour of the day.
- 1China—
- 2United States—
- 3Japan—
- 4South Korea—
- 5Brazil—
- 6Germany—
That works out to roughly one Photoshop job every 12 seconds.
And they are not just connecting. In the same two weeks the server recorded on the order of 100,000 tool batches — real Photoshop work, not empty handshakes.
I did not want this to become another cloud editor. Photoshop stays on your machine. The agent talks to it locally. That is a product decision as much as a technical one: creative files are the last thing most teams want uploaded “for convenience.”
You do not have to wire Cursor to try it. The package now ships a local chat window, because the people who needed this were not all developers:

npx -p @alisaitteke/photoshop-mcp photoshop-mcp-ui
Sign in with an API key, or reuse a Claude Code / Gemini CLI account.
The design decisions that actually mattered
The first version exposed a lot of Photoshop as MCP tools and hoped the model would compose them. That is the demo that looks good in a README. It is also how you get brittle scripts, half-applied filters, and a document nobody wants to undo through.
So I changed the product surface.
Recipes, not just API coverage. A recipe is one outcome — remove the background, enhance a portrait, export for web — and one Undo in Photoshop. The LLM still has 102 atomic tools when it needs precision. For the jobs people actually ask for, it should not have to rediscover a twelve-step procedure every time.
State before action. Agents guess. Photoshop does not forgive guesses. The useful loop is: read capabilities, read document state, act, then look at a preview. Structured errors on the way back, so the next attempt is not the same prompt shouted louder.
Plan once, execute once. Action Plan (beta) is the same instinct. Multi-step prompts used to mean a round-trip per click. Now the model can plan the whole pass, then run it. Fewer tokens, fewer half-finished documents.
Meet people where they already work. Cursor and Claude for people who live in an IDE. A standalone UI for everyone else. Cross-platform because Photoshop users are not a single OS: AppleScript on macOS, COM on Windows, versions that go back further than most agent demos bother with.
None of that is glamorous. It is the difference between a protocol wrapper and a tool a stranger will leave running.
What changed since February
Fifty tools became a product:
- 118 tools — 102 atomic
photoshop_*operations plus 16 recipes. - Standalone web UI — chat without an IDE.
- Action Plan (beta) — plan every step in one LLM call, then run them in a single pass.
- A proper site at photoshop-mcp.com.
The recipes are the part I wish had existed on day one. You say what you want; the server already knows the steps:
Remove the background from this portrait — keep it editable with a mask.
Enhance this portrait — smooth the skin and fix the tones, medium intensity.
Prepare this design for web, then export Instagram and X post variants.
Split this wide banner into a 5-slide seamless Instagram carousel.
What it is not
It is not an Adobe product. It is not a SaaS that hosts your PSDs. It is not “AI Photoshop in the browser.”
The promise is smaller, and I think more useful: if Photoshop is already on the desk, an agent should be able to operate it — locally, undoably, without you babysitting every click.
That is the class of problem I like working on. Not a new surface for its own sake. The glue between models and the software organisations already standardised on.
If you want to try it
Photoshop running, Node.js 18+, Windows or macOS.
Chat window (easiest):
npx -p @alisaitteke/photoshop-mcp photoshop-mcp-ui
Inside Cursor, Claude Desktop, or VS Code — add to MCP config:
{
"mcpServers": {
"photoshop": {
"command": "npx",
"args": ["-y", "@alisaitteke/photoshop-mcp"]
}
}
}
Claude Code:
claude mcp add photoshop -- npx -y @alisaitteke/photoshop-mcp
The February write-up is still there if you want the original pitch: Photoshop MCP Server.
Links
Photoshop MCP is MIT licensed.
If you are wiring agents into tools your team already depends on — creative, internal, or otherwise — I hope this is a useful data point. I am still learning what “reliable enough” looks like when the other end of the socket is a 30-year-old desktop app, so feedback and strange edge cases are very welcome.
