GMI MCP Server: turn your assistant into a production studio

A landing page needs a hero image. A highlight reel needs a track under the b-roll. A keynote needs a ten-second title card. A tutorial script needs a voice reading it. A launch post needs five product shots in different settings. A support reply needs the exact name of a parameter buried three pages into a doc.

In every case, the media or the answer is one step inside something bigger: the project. The GMI MCP Server puts that step inside the tool where the bigger thing is already happening. Connect it to Claude, Cursor, ChatGPT, or Codex, and ask for what you need in the same sentence you'd use to describe it to a coworker.

What builders are asking for

How one of these plays out

Say you're editing a product demo and the last shot needs a five second clip of a still photo coming to life with a slow push in. You type the request into Claude Code the way you'd describe it to an editor: turn this photo into a five second video with a slow camera push in, and tell me what it costs before you run it.

Your assistant reads the model's parameters instead of guessing at them, since video models each name settings their own way. It checks the exact resolution and duration you asked for, quotes a price, and waits. You approve. Because a video job takes minutes rather than seconds, you get a tracking reference right away and a note confirming the job is actively running. Ask again in a few minutes and the finished clip comes back, ready to cut into the sequence you're already working in.

That five-step shape, describe it, read the parameters, price it, approve it, collect it, is the same whether you're asking for a video, an image, a voiceover, or a track. Only the parameters and the pricing unit change.

But, What is MCP?

MCP, or Model Context Protocol, is a standard that lets AI assistants like Claude, ChatGPT, or Cursor connect directly to outside tools and services. Instead of an AI only knowing what you type, MCP gives it a way to reach out, check real data, and take real actions, like generating an image, checking an account balance, or looking something up, all through plain conversation instead of custom code.

API Vs MCP

We ran the same request, a man in a park with a dog, two ways: once by hand against GMI's raw API, once through the GMI MCP Server. The image that came back was equally strong on both paths. The difference showed up entirely in how many steps it took to get there.

Both paths produce usable results. The raw API gives a backend full control over every parameter and retry. The GMI MCP Server gives an assistant the same access with the setup already done, which is the difference that matters most in the middle of building something else.

An example: turning a script into narration

A different shape, same pattern. You've finished a five minute tutorial script and need it read aloud for the screen recording you're about to cut together. You ask your assistant what voice options exist, pick one that sounds clear and neutral, and ask for the script read back at that voice. Since text-to-speech is priced per character rather than per second, your assistant quotes the exact cost for the script length you gave it. You approve, the audio comes back, and you drop it under the recording, no separate narration tool or manual export step involved.

The pattern holds for a team checking in on spend, too. Someone on your team asks what the account spent on generation over the past week and which models it went to. The answer comes back broken down by model, over the exact window asked for, so a quick question during a standup gets answered on the spot.

Why this sits inside the chat instead of a separate tool

Generating a piece of media usually means leaving the page you're writing, the deck you're building, or the edit you're cutting, opening a different tool, describing the same thing again in that tool's own interface, and bringing the result back yourself. The GMI MCP Server removes the middle step. You describe what you need once, where you already are, and your assistant handles picking a model, shaping the request, and returning a finished file or a link to one.

What's available

The server covers four kinds of requests, and your assistant picks between them based on what you ask for:

Each generation request is priced and shown to you before it runs, so approving a job means you already know what it costs. Generating and charging both happen only after you approve.

More tools are on the way, expanding beyond images, video, and audio into new parts of the GMI Cloud platform. Each one will follow the same pattern: ask in plain language, see what it involves, and approve before anything runs.

Connected everywhere you already work

The server lives at one address, and every supported client points at the same one: Claude, Claude Code, Codex, ChatGPT, and Cursor all connect the same way, through a one-time browser sign-in. Sign-in happens once through the browser, and one address covers setup for every client. Whichever tool you're already writing in, the same set of requests works the same way inside it.

Built for how small teams work

This fits the team you already have, generation as a side task rather than a dedicated job. A two person team shipping a landing page, a solo developer recording a course, a marketer cutting a highlight reel between meetings, all of these are the same shape of request: something is almost done, and one piece of media or one quick answer is what's left. The GMI MCP Server is built around that moment rather than around a production workflow with its own tooling and its own login.

That also means the same server scales down cleanly. A single request costs a single request. The first image or clip comes back on the same login and plan you already have.

Getting started

Point your client (Claude, Claude Code, Codex, ChatGPT, or Cursor) at GMI MCP, sign in once through your browser, and you're connected.

In Claude Code that's one line in the terminal; in Cursor, Claude Desktop, or any JSON-configured client, it's one block in the MCP settings file.

claude mcp add --transport http gmi https://mcp.gmicloud.ai/mcp

{
  "mcpServers": {
    "gmi": {
      "url": "https://mcp.gmicloud.ai/mcp"
    }
  }
}

Full setup steps for each client are in the GMI MCP Server docs. 

Once it's connected, ask for the next piece your project needs.