Relay

What the Relay MCP server does

The Relay MCP server gives an AI client access to Relay’s model catalog, model execution API, workflow system, file hosting, documentation, and account reporting. Models are selected by name, so newly available Relay models can be used without updating the MCP package. Supported model types include image, video, audio, and language models.

Agents can search models by text or tag, inspect a model’s route, input schema, example, price, and status, and query public pricing. They can also estimate a model’s cost from the published price list. Execution tools accept model-specific inputs and wait for a bounded period before returning a result or task identifier.

How it works

Relay provides two connection modes. The hosted endpoint uses Streamable HTTP at https://mcp.relaygpu.com/mcp; clients send a Relay key through the X-API-Key header or an Authorization Bearer token. Nothing needs to be installed for this mode. The local option starts npx -y @relaygpu/mcp as a stdio process and reads RELAY_API_KEY from its environment. Its default API base URL is https://relaygpu.com.

Short operations can wait for completion directly. Model and workflow calls wait for up to 45 seconds, while task and workflow polling can wait for up to 30 seconds. If work continues, the agent can poll it later, and an in-progress workflow can be cancelled. The hosted process reads the credential from each request; the local process reads it from its environment, and the documentation states that keys are not stored or returned in tool results and logs.

Setup and configuration

For hosted connections, configure the endpoint and credential in the client. The README documents setup for Claude Code, Codex, Cursor, VS Code, and Claude Desktop. Claude Desktop uses the downloadable relay.mcpb bundle for the remote-service limitation described in the README, while other clients can use their documented HTTP configuration. A local setup requires Node 20 or newer and the @relaygpu/mcp package through npx.

Set RELAY_API_KEY for local stdio execution. RELAY_BASE_URL can override the default Relay API URL. In the hosted configuration, use either an API key header or a Bearer token. get_credits and get_usage specifically require a dashboard JWT or custom-tier superkey rather than a plain API key.

Tools and capabilities

Available capabilities include:

• Find models with search_models, inspect one with get_model, and query prices with get_pricing.
• Estimate public-list costs with estimate_cost.
• Run models with run_model, then inspect unfinished work with check_task.
• Upload local files using upload_file to obtain URLs for model inputs ending in _url.
• List workflow templates, start runs, poll them, and cancel runs that remain in progress.
• Search Relay documentation with search_docs.
• Read credit balances, promotions, consumption, and usage analytics with the account tools.

Discovery, pricing, task-status, and documentation tools marked keyless in the README do not require a credential. Model execution, uploads, workflows, cancellation, and account tools require authentication as applicable.

Limitations and notes

The Relay MCP server does not make every long-running operation synchronous: calls return after their wait limit if the job has not finished, so the agent must use the relevant polling tool. Public cost estimates are estimates from the price list, not invoices. A missing credential produces a tool error identifying the required authentication setting, and Relay errors include an error class, code, request ID, and server detail.

Local startup depends on the client being able to find npx. Applications launched outside a shell may not inherit a Node installation managed by tools such as nvm, Volta, or Homebrew. Windows users may need to invoke npx through cmd. The first npx launch downloads the package and can take several seconds; later starts are faster.