Manual tool walkthrough

Use this guide when you are integrating Scrollport directly and need the explicit control-tool calls. If your agent is already connected, the quickstart is the simpler path: describe the outcome and let the agent handle this sequence.

The catalog response and tool input are live, so placeholders below stand for values returned by the current API rather than a claim that a particular provider call ran.

Start safely

Give the agent the recommended setup prompt, complete the browser approval and call get_wallet. Confirm available is enough for the selected job before asking it to run.

Discover the outcome

{ "name": "search_tools", "arguments": { "query": "find the text of a web page" } }

Select a tool_id from the response. The provider name is provenance, not the job to search for.

Inspect the selected tool

{ "name": "inspect_tool", "arguments": { "tool_id": "<returned tool_id>" } }

Read input_schema, worked_example and common_errors. Copy the worked input and substitute your own value without changing its shape.

Run once

{
  "name": "run_tool",
  "arguments": {
    "tool_id": "<returned tool_id>",
    "input": { "<inspected field>": "<your value>" },
    "idempotency_key": "<one client-generated UUID for this logical start>",
    "wait_seconds": 0
  }
}

If the response is still queued or running, keep the run_id and read it with get_run:

{ "name": "get_run", "arguments": { "run_id": "<returned run_id>", "wait_seconds": 50 } }

This read needs no idempotency key and cannot start another run. If the run asks for approval, stop and show the human the estimate, then use get_run after approval. If it asks for a connection, give the human the connection URL.

Reconcile the wallet

Call get_wallet after the run. A successful run has cost_final; a failed run has released its hold and is not billed. This final read makes the money result visible rather than treating provider output as the only success signal.