FAQ

Short answers on actions — what runs without asking, what the visitor can see, how failures show, and what an MCP server needs to work with Deskbell.

Does anything run without me choosing it?

No. Every discovered tool is off until an admin sets a policy, and the suggestion is Allowed only for tools the server marks read-only.

Can the visitor see what a tool returned?

Anything a tool returns may end up in the answer. The model is told to use results as material and to cite them, but the text reaches the model and the model writes the answer. Attach a server the way you would give a support agent access to it, and use the identity token to scope what it returns.

What does a failed call look like?

The step line says Couldn't complete, the model is told the call failed and answers without it, and the audit row records the error. A call that takes longer than the timeout counts as failed.

Can a tool call another tool?

Within a turn the model may call several tools, up to the step limit. A tool that needs confirmation pauses the turn; after the decision the model may continue calling allowed tools.

What does my MCP server need?

A remote endpoint over Streamable HTTP (or SSE), reachable from the Deskbell instance over https, a tools/list that describes each tool with a JSON Schema input, and either no credential, a bearer token or a header value. Annotate read-only tools with readOnlyHint so the suggested policy is right. To scope results to the visitor, verify the identity token.

Does the eval harness cover answers that used tools?

The harness replays questions against the content; it declares no tools, so those runs are unchanged. Answers that used a tool are judged the way any answer is when they appear in real traffic and are exported as cases.

Where is the record?

One row per attempted call in the action_calls table: input, output (cut at 16 000 characters), status, the visitor it ran for, the decision and its time, latency and error. The Conversations page shows the calls per answer.