MCP server

Connect your agent to one MCP server. Now it can do anything.

Your coding agent can write code. With one InferAll connection it can also ask Claude, GPT or Gemini for a second opinion, generate an image, render a video, and read a script out loud. One key, one line of setup, no account per provider.

Get a free key@inferall/mcp on npm

Setup in one line

Create a key at /keys. Put it in your shell as INFERALL_API_KEY, for example in ~/.zshrc, not in a file you commit. Then add the server.

Claude Code

claude mcp add inferall -e INFERALL_API_KEY="$INFERALL_API_KEY" -- npx -y @inferall/mcp

Cursor, Claude Desktop, Windsurf and other mcpServers clients

{
  "mcpServers": {
    "inferall": {
      "command": "npx",
      "args": ["-y", "@inferall/mcp"],
      "env": { "INFERALL_API_KEY": "ifu_..." }
    }
  }
}

Six tools

  • chat: Prompt any chat model: Claude, GPT, Gemini, Grok or the open NVIDIA NIM models. The default is a $0 NIM model.
  • generate_image: Text to image with FLUX or Gemini image models. Returns image URLs.
  • generate_video: Text or image to video with Runway, Seedance, Veo, Kling, Hailuo or Wan. Returns the video URL.
  • text_to_speech: ElevenLabs speech, saved as an MP3 file on your machine.
  • list_models: Search the live model list by a word such as video, image or claude.
  • account_status: Free trial calls left, balance, and whether paid models are unlocked.

What you can ask for

  • “Ask Claude and Gemini to review this diff and tell me where they disagree.”
  • “Make a hero image for this landing page and save the URL in the README.”
  • “Render a five second product video of the screenshot in ./demo.png.”
  • “Read the release notes out loud and save the MP3 next to them.”
  • “Which video models can I use, and how many free calls do I have left?”

What it costs

The server is free. The open NVIDIA NIM chat models are free on the trial. Paid chat models, images, video and speech draw from credit at the rates on /models. Add $5 on /billing to unlock them.

A refused call comes back as a tool error with the reason: an unknown model id, a paid model on a trial key, or a used-up trial. The server never swaps in a different model without telling you.

Questions

What is the InferAll MCP server?

It is an open npm package, @inferall/mcp, that gives any MCP client six tools: chat, generate_image, generate_video, text_to_speech, list_models and account_status. Every tool calls the InferAll API with your one key, so the agent reaches chat, image, video and speech models without a separate account per provider.

Which agents does it work with?

Any client that runs a local stdio MCP server. That includes Claude Code, Cursor, Claude Desktop and Windsurf. Clients that take an mcpServers JSON block use the config shown on this page.

What does it cost?

The open NVIDIA NIM chat models are free on the trial. Paid chat models, images, video and speech draw from credit at the rates on /models. A $5 credit pack on /billing unlocks them.

What happens when a call is refused?

The tool returns an error with InferAll's own message, for example an unknown model id, a paid model on a trial key with the billing link, or a used-up trial. The server never swaps in a different model without telling you.

Why does a video call take minutes?

Video models usually take 1 to 6 minutes to render. The tool waits for the result. If your client stops tool calls sooner, raise its limit; in Claude Code set MCP_TOOL_TIMEOUT=600000.