MCP

Transport

Streamable HTTP, stateless: every request stands alone and is answered with JSON; no session, no long-lived stream.

Auth

The same key as the API, in a header of the MCP connection: Authorization: Bearer sk-qovia-… (or x-api-key). Searching and reading endpoints needs no key; calls and the account check do.

Put your own key in the config below, restart the client, and qovia shows up among its tools.

Claude Code

claude mcp add --transport http qovia https://api.qovia.ai/mcp \
  --header "Authorization: Bearer $QOVIA_KEY"

Cursor

{
  "mcpServers": {
    "qovia": {
      "url": "https://api.qovia.ai/mcp",
      "headers": {
        "Authorization": "Bearer sk-qovia-…"
      }
    }
  }
}

VS Code

The key is asked for on first use and never written to the file:

{
  "servers": {
    "qovia": {
      "type": "http",
      "url": "https://api.qovia.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${input:qovia-key}"
      }
    }
  },
  "inputs": [
    {
      "type": "promptString",
      "id": "qovia-key",
      "description": "Qovia API key",
      "password": true
    }
  ]
}

Claude Desktop

Claude Desktop's remote connectors use OAuth, so bridge with mcp-remote (needs Node.js):

{
  "mcpServers": {
    "qovia": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://api.qovia.ai/mcp",
        "--header",
        "Authorization:${QOVIA_AUTH}"
      ],
      "env": {
        "QOVIA_AUTH": "Bearer sk-qovia-…"
      }
    }
  }
}

By default the AI gets four tools, so the catalog never crowds its context: find, read the parameters, then call.

To give the AI the endpoints themselves, add ?tools=all, or only some platforms:

Every live endpoint then becomes a tool named after its id with dots as underscores (raw.douyin.video → raw_douyin_video), taking the endpoint's own parameters. The four default tools stay.

Given a data question, the AI usually goes like this:

User: "How many likes and comments does Douyin video 7681194457378688292 have?"
→ search_endpoints {"query": "douyin video"}
→ get_endpoint {"endpoint_id": "raw.douyin.video"}
→ call_endpoint {"endpoint_id": "raw.douyin.video", "params": {"item_id": "7681194457378688292"}}
← data.metrics.likes / comments, meta.charged_usd = 0.0015

• Every MCP call runs the same pipeline as the API: auth, scopes, parameter checks, rate limits, billing and the usage log are identical.
• Only successful calls that return data are charged; empty results and failures are free. Free quota applies first where an endpoint allows it.
• The per-minute limit is per account, shared with the API and the console playground.
• Errors come back as tool errors with error.code (the API's error codes) and the HTTP status, plus retry_after_seconds when rate limited.