MCP for Claude, Cursor and ChatGPT

https://ugesi.com/api/mcp/

Authorization: Bearer ugsk_your_key

https://ugesi.com/api/mcp/ugsk_your_key/

{
  "mcpServers": {
    "ugesi": {
      "command": "node",
      "args": ["/path/to/ugesi-mcp.mjs"],
      "env": { "UGESI_API_KEY": "ugsk_your_key" }
    }
  }
}

claude mcp add --transport stdio --env UGESI_API_KEY=ugsk_your_key ugesi -- node /path/to/ugesi-mcp.mjs

{
  "mcpServers": {
    "ugesi": {
      "command": "node",
      "args": ["/path/to/ugesi-mcp.mjs"],
      "env": { "UGESI_API_KEY": "ugsk_your_key" }
    }
  }
}

[mcp_servers.ugesi]
command = "node"
args = ["/path/to/ugesi-mcp.mjs"]
env = { "UGESI_API_KEY" = "ugsk_your_key" }

https://ugesi.com/api/openapi/

{
  "mcpServers": {
    "ugesi": {
      "command": "node",
      "args": ["/path/to/ugesi-mcp.mjs"],
      "env": { "UGESI_API_KEY": "ugsk_your_key" }
    }
  }
}

Generate without leaving the assistant

Ask Claude for a product shot on marble. It calls UGESI, waits for the job and hands back the file. No tab switching, no copying prompts.

Five tools

List models with rates. Estimate a job before running it. Start a generation. Check a job. Compose a recipe from guided mode, the same vocabulary the site uses.

Estimate before you generate

Ask the assistant what a job costs and it calls the estimate tool. Credits, a dollar figure and whether the model is available. Nothing is held and no balance is needed, so planning a batch costs nothing.

Billing

Every call draws from the same credit balance as the browser. Failed jobs refund themselves. Paused or retired models return a clear error with alternatives.

Who can use it

Anyone with a credit pack bought in the last 12 months. Agents generate in loops, so every key gets a rate limit, a usage view and a way to revoke it at once.

What an assistant can and cannot do here

It can list models with current rates, start a generation, check a job and read a recipe. It cannot spend beyond your balance, change your account or access anything outside the five tools. The key you issue defines the limit.

Why this is not just an API wrapper

An assistant that can only generate has to guess which model to use. These tools expose rates and limits first, so the assistant can reason about cost before spending your credits. That changes the quality of its decisions.

What it costs to run

The same rates as the interface. An assistant spending your credits is still your balance. Every job appears in your history with the model, the prompt and the cost.

Where it is genuinely useful

Batch work with judgment in it. Twenty product shots, each on a different background. Or a set of variants where the assistant checks the result and reruns the ones that missed. That is work a form cannot do and a script does badly.

Keeping an assistant inside a budget

Give the assistant its own key. Its usage shows separately in your account, and revoking it stops the assistant without touching other integrations. Runs draw from the account balance, so the balance is the upper limit.

What to ask for

Specific jobs work better than open ones. Ask for six product shots on different backgrounds and it will plan the runs. Ask it to make something nice and it will spend your credits finding out what that means.

Reading the history afterwards

Every job an assistant runs appears in your history with the model, the prompt and the cost. Exactly like a job you started yourself. There is no separate log to reconcile and nothing hidden behind an agent.

What to do when a tool call fails

The five tools return the same errors as the API, in plain language the assistant can act on. Not enough credits names the amount. A paused model names replacements. An assistant that reads the error can usually fix the request itself.