No tool call is a valid answer
When nothing fits, the most useful thing an engine can do is decline. Bluwing returns no tool call instead of reaching for the closest one.
Ask a tool-calling system what's the capital of Peru and watch what happens. If every response must contain a tool call, something has to be chosen. The model picks the least-bad option, your app runs a function that has nothing to do with the question, and the trace looks successful.
Bluwing returns no_tool. It is one of the most useful things the engine does.
Why the wrong tool is worse than no tool
A missing call is visible. Your app sees that nothing matched and can answer in its own words, ask a clarifying question, or fall back to a plain model. A wrong call is invisible until it has side effects.
Consider a support agent with refund_order, cancel_order and lookup_order. A vague message — this needs sorting out — has no safe default. Choosing cancel_order because the word order appears is not a guess the system should make on the user's behalf.
How the engine decides
The narrowing step produces a ranked branch. A call is only emitted when:
- a branch clears the relevance bar;
- a leaf inside it clears the tool-level bar;
- all required arguments are grounded.
If any of those fails, the turn ends without a call. The response still has the standard shape, so your client does not need a special code path:
{
"finish_reason": "no_tool",
"tool_calls": []
}That is the whole contract. Your app can treat it exactly like a model reply with no tool use.
What your app should do with it
There is no single right answer, but there are three good ones:
- Answer directly. If your product has a plain chat path, route the turn there.
- Ask. I can look up orders or start a refund — which one do you need?
- Do nothing. In a narrow workflow, declining is a correct outcome.
What you should not do is retry with a broader tool list until something matches. That turns a clean decline into the forced guess the engine was avoiding.
The other side: needs_input
No tool and missing input are different outcomes, and the engine keeps them separate.
| Status | Meaning | What happens next |
|---|---|---|
tool_calls | A tool was chosen and its arguments are grounded | Run the call |
needs_input | A tool was chosen but a required value is missing | Ask for the named field |
no_tool | Nothing fit | Your app decides |
Mixing the last two is a common bug. A missing party size does not mean "no tool fits"; it means the booking tool is right and one field is absent. Keeping them distinct lets your UI respond correctly in both cases.
The pressure to always answer
Product teams dislike empty states, and it is easy to add a fallback "just pick something" mode. Resist it. The value of a tool-calling layer is that it knows what it can and cannot do. An engine that always answers is an engine whose answers you cannot trust.
If you need coverage for a question the tools cannot serve, add a tool for it or route it to a model. Do not widen the definition of a match until everything matches.
Frequently asked questions
Does no_tool cost a credit?
No. You pay one credit per successful resolve. A turn that returns no call is free.
Can I force a tool call?
Not by default. Forcing a call is how invented arguments and wrong tools get through. If your workflow requires an action, make the request more specific or ask the user.
How is this different from a low-confidence answer?
There is no confidence score to tune. The engine either clears the bar or it does not. A threshold you can slide is a threshold that eventually gets slid for the wrong reason.