MCP server

Connect Claude Desktop, Cursor, or any Model Context Protocol client to Katama and generate images & videos straight from your assistant — metered against your Katama credits.

Endpoint

The MCP server speaks JSON-RPC 2.0 over Streamable HTTP at:

https://katama.ai/api/mcp

1. Create an API key

Standard MCP clients authenticate with a bearer token. Create one in Settings → API keys (MCP / API access). The full key (kata_sk_…) is shown once — copy it immediately; only a hash is stored, so it can never be retrieved again. You can revoke a key at any time.

2. Add it to Claude Desktop

Add a katama entry to the mcpServers block of your Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "katama": {
      "type": "http",
      "url": "https://katama.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer kata_sk_your_key_here"
      }
    }
  }
}

Clients that launch MCP servers via a local bridge can use mcp-remote instead:

{
  "mcpServers": {
    "katama": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "https://katama.ai/api/mcp",
        "--header", "Authorization: Bearer kata_sk_your_key_here"
      ]
    }
  }
}

Tools

Tool	Cost	What it does
`list_models`	Free	Lists chat + image/video models, their credit cost, and which your plan allows.
`get_balance`	Free	Your current Katama credit balance.
`enhance_prompt`	Free	Improves an image/video/text prompt for better results.
`chat`	Credits (flat, 1–7/message)	Chat with Katama's LLMs — Claude Opus/Sonnet/Haiku, GPT 5.5, Gemini, DeepSeek, GLM, MiniMax, Kimi, Grok. Default model `auto` picks the cheapest-good-enough.
`generate_image`	Credits	Generates an image from a prompt. Returns the image URL + credits charged.
`generate_video`	Credits	Generates a video from a prompt (optionally from a start image).

Billing & safety

Credits are reserved before generation and released if it fails — a failed or unconfigured generation never costs you anything.
The cost is computed server-side from the model and clamped parameters, so the estimate always equals the charge. Prices are never taken from the client.
Every prompt passes the same content-safety guard as the web app.
Plan-gated models require the matching plan; list_models shows what your plan allows.