Testing in the sandbox

The sandbox lets you run one of your server's tools from the app, without an AI, to check that it does what you expect.

Choose a tool

Open the server's "Test" tab and type part of a tool's name or description in the "Tool" field. Several words work in any order, ignoring case and accents. Tools whose name matches appear first. If nothing matches, "No tools found" is shown and the current tool does not change. With the keyboard, use the up and down arrows to move, Enter to choose, Escape to close the list (press it again to clear the field) and Tab to leave without choosing. Only active tools are offered. Read-only tools are marked "Read only"; the others may change or delete data in your target API, so test them with care. If the list is empty, import an OpenAPI spec or add documents in the "Sources" tab.

Enter arguments with the form

"Form" mode shows one field per parameter, with a suitable control (number, integer, boolean "Not set" / true / false). Required fields are flagged: "Required.".

Enter arguments as JSON

"JSON" mode opens an "Arguments (JSON)" editor. The arguments must be a JSON object, for example {"city": "Paris"}. Invalid JSON, or JSON that is not an object, is flagged before sending.

Read the result

Click "Run". The "Result" block shows:

• "Success" or "Tool error" (the target API answered with an error status, or the call failed);
• the call "Duration";
• the "HTTP code" returned by the target API, when there is one;
• the response, as the AI would receive it (see Target API).

Quota

Each run counts toward your monthly quota, like an AI call, and appears in the call log. Only invalid arguments (rejected before the call) are neither counted nor logged. When the quota is reached, the sandbox returns the QUOTA_EXCEEDED error (details) and does not call your API. See Quotas and plans.

Inactive server

The sandbox works even if the server is not "active": handy to test before connecting an AI. Connected AIs, however, are refused (403) until the server is active.

OAuth servers

On a server whose target API uses OAuth 2.0, the API tools need the token of a person signed in at the provider, which only an AI connected with OAuth has. The "Test" tab lists them, but "Run" is disabled; a call made anyway is refused with OAUTH_REQUIRED, without counting toward the quota. Document tools (search_documents, read_document) stay testable here. To try an API tool, connect your AI and ask it to call the tool.

Forwarded credential

On a server whose target API uses "Forwarded by caller", the "Test" tab shows a masked "Forwarded credential" field for the API tools. Enter the value your API expects; it is sent for this run only and never stored. See Per-user credentials with an OAuth gateway.

Example

Your "Weather" server is still a "draft". In "Test", you pick the get_forecast tool, switch to "JSON" mode and enter {"city": "Paris"}. After "Run", the result shows "Success", a duration of 340 ms and HTTP code 200 with the forecast. You know the tool is ready before you connect your AI.