WaveSpeed MCP

OfficialRemoteby WaveSpeed AIcreditsverified Oct 19, 2026

WaveSpeed's live image, video, audio and 3D catalogue behind seven tools, hosted or via npx.

What it does

WaveSpeed's own server reads its model catalogue from the live API instead of a pinned list, so an agent searches with list_models, reads a model's real input schema with get_model_schema, quotes the job with get_price, then fires run_model. For a filmmaker-agent that is one bill and one verb across many upstream image and video models, with an explicit price check before an expensive shot and resumable polling for renders that outlast a single tool call.

Connect

Remote endpoint (Streamable HTTP)

https://api.wavespeed.ai/mcp

Claude Code

claude mcp add --transport http wavespeed https://api.wavespeed.ai/mcp

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "wavespeed": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://api.wavespeed.ai/mcp"
      ]
    }
  }
}

Cursor (.cursor/mcp.json)

{
  "mcpServers": {
    "wavespeed": {
      "url": "https://api.wavespeed.ai/mcp"
    }
  }
}

Auth: oauth · env WAVESPEED_API_KEY — https://wavespeed.ai/accesskey — the hosted server signs in through the browser and mints a key named 'MCP · <app>'; the npx server and header-only clients take WAVESPEED_API_KEY or an Authorization: Bearer header

Sample tools

list_modelsget_model_schemarun_modelget_priceget_balanceupload_fileget_prediction

Field notes

Verified 19 Oct 2026 from the repo, npm and wavespeed.ai/docs/mcp. MIT, 34 stars, repo pushed and npm @wavespeed/mcp 1.0.11 published that same day; the package has existed only since 19 Aug 2026, so it is young and moving fast. api.wavespeed.ai/mcp answers 405 to a plain GET, as a streamable-HTTP endpoint should. The good: get_price discloses which inputs its quote ignored (unpriced_inputs) instead of passing off a floor as the price, and run_model rejects input keys the model does not accept before you are charged. The caveats: the vendor says the quote is an estimate and the amount actually charged is what counts; the hosted server cannot read local files, so pass public URLs or use its create_upload flow, while @path uploads work only in the npx version; and on the hosted server run_model waits at most 100 seconds (1,800 locally) before handing back a prediction id to poll with get_prediction. Deleting the 'MCP · <app>' key disconnects a client. A separate community Python package called WavespeedMCP exists; this entry is the vendor's own. The lane is crowded — compare fal, replicate, runware and atlas-cloud on catalogue and price.