MCP Server

What the gpu-bridge/mcp-server MCP server does

gpu-bridge/mcp-server MCP server connects MCP-compatible clients to GPU-Bridge's catalog of 30 hosted AI services. The catalog includes language-model inference, code generation, image generation and editing, video generation and enhancement, speech transcription, text-to-speech, music and sound-effect generation, embeddings, document reranking, OCR, PDF parsing, image analysis, background removal, and content moderation.

The main execution tool accepts a service identifier, service-specific input, and an optional priority selection. Available priorities are fast for lower latency or cheap for lower cost. The service catalog and cost estimator can be used before submitting an inference request, which helps an agent select a model and estimate usage.

How it works

The gpu-bridge/mcp-server MCP server exposes five tools:

• gpu_run submits work to a selected GPU-Bridge service.
• gpu_catalog lists available services, models, capabilities, and pricing.
• gpu_estimate estimates the cost of a service request without authentication.
• gpu_status checks a job and retrieves its result.
• gpu_balance reports credit balance, daily spending, volume-discount tier, and job history.

GPU-Bridge supports two payment paths. A client can send an API key, or an autonomous agent can use x402 to pay per request with USDC on Base L2. With x402, the provider returns an HTTP 402 payment response, the agent submits payment, and the request is retried with payment proof. The README describes this flow for direct API use; the MCP configuration example uses an API key.

Setup and configuration

Install the package from npm and register it as a local stdio server. For Claude Desktop, add an entry under mcpServers and restart the application:

{
  "mcpServers": {
    "gpu-bridge": {
      "command": "npx",
      "args": ["-y", "@gpu-bridge/mcp-server"],
      "env": {
        "GPUBRIDGE_API_KEY": "your_api_key_here"
      }
    }
  }
}

The API key is obtained from GPU-Bridge. The README also documents a keyless x402 option for agents that can make USDC payments, but it does not provide a separate MCP configuration example for that mode. Claude Desktop configuration paths are listed for both macOS and Windows.

Tools and capabilities

Use gpu_catalog to inspect current service keys rather than assuming a model is available. Examples in the catalog include llm-4090, flux-schnell, whisper-l4, tts-l4, embed-l4, rerank, animatediff, video-enhance, and pdf-parse.

The catalog covers sub-second LLM options, image generation, multilingual OCR, visual question answering, speaker diarization, voice cloning TTS, music generation, code embeddings, document reranking, and video upscaling up to 4K. Actual inputs vary by service, so callers should consult the service information before constructing a gpu_run request.

Limitations and notes

Inference is pay-per-use rather than described as a fully local operation. Listed starting prices include separate rates for language-model tokens, images, transcription minutes, speech characters, embeddings, reranking queries, and document parsing. Prices and service availability should be checked through gpu_catalog or GPU-Bridge's published catalog.

The README explicitly shows Claude Desktop and states that other MCP-compatible AI clients can use the server, but it does not document client-specific setup for Cursor, Windsurf, or Cline. The server's listed tools cover GPU-Bridge services; they do not expose arbitrary local GPU workloads.