Set up with MCP

NativePort speaks the Model Context Protocol over Streamable HTTP at https://api.nativeport.ai/mcp. Your agent connects in one of two ways: it signs in with your NativePort account in the browser, or it sends an API key in the Authorization header. Either way it gets four tools that mirror the Agent API: find an endpoint, read its price, call it, check the balance. No SDK, no skill file, no per-provider setup.

Sign in with your account (recommended)

A client with a sign-in flow for MCP servers connects without a key; Claude, Claude Code, ChatGPT and Codex already do. No key is copied or pasted, so none can end up in a chat or a config file.

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

Calls from a signed-in app are billed to your account’s balance, exactly like calls made with a key.

Connected apps

Every app you approve appears under Connected apps at panel.nativeport.ai/keys. Revoke cuts it off on its next request, even while its sign-in is still valid.

If your account is not set up yet

Signing up on the client’s sign-in page, without ever opening the panel, leaves the account unfinished. The server still connects and find_endpoints and inspect_endpoint work, but account and call_endpoint answer:

Your NativePort account is not set up yet. Sign in once at https://panel.nativeport.ai, then try again.

Open panel.nativeport.ai once with that account. The next call works, and the client does not need to reconnect.

Clients that cannot sign in

The sign-in identifies each client by a metadata document it publishes at a URL (Client ID Metadata Documents). A client that only supports dynamic client registration cannot sign in yet; connect it with an API key.

Use an API key (no browser)

For servers, CI and agents without a browser, send your key in the Authorization header.

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

{
  "mcpServers": {
    "nativeport": {
      "url": "https://api.nativeport.ai/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_KEY" }
    }
  }
}

Tools

The resource nativeport://skill returns the same SKILL.md the gateway serves over HTTP.

call_endpoint runs the ordinary gateway route internally, so billing, logging and every restriction apply exactly as for a direct HTTP call. Only your Authorization header and the body’s content-type are forwarded; other request headers cannot be set through MCP.

A few routes take a form instead of JSON: ElevenLabs speech-to-text answers a JSON body with 422. Pass their fields in form (strings, numbers or booleans), and call_endpoint sends them as multipart/form-data. body and form cannot be combined. Uploading a file still needs a direct HTTP call.

What comes back

• Responses are read for a model: JSON is parsed, text is capped at 512 KB (truncated: true when cut), a JSON array over the cap comes back as its first complete items with truncated: true and returnedItems (page the rest with the endpoint’s own limit/offset), a JSON object over the cap is cut down to fit when it can be, with truncated: true and each cut listed in omitted with its path and size (a list keeps its first items whole and counts the rest in omittedItems, a string keeps its start and ends in …[cut], and anything else is left out as [left out: N bytes]), an object over 8 MB or one that cannot be cut down comes back as cut text, binary bodies under the cap come back base64-encoded with their MIME type and larger ones are omitted with a note. Do not request streaming through call_endpoint.
• A non-2xx answer is a tool result with isError: true and the gateway’s body verbatim. 402 means the balance is empty and the agent should stop and tell you; 403 with "Account suspended.", a trial-restriction body or the account-not-set-up sentence should be reported as is; 404 on a documented path means the action is not admitted for your account.

Connection semantics

The server is stateless: there are no sessions to resume, and each request stands alone.