Trait Tool 

pub trait Tool:
    Send
    + Sync
    + Attributable {
    // Required methods
    fn name(&self) -> &str;
    fn description(&self) -> &str;
    fn parameters_schema(&self) -> Value;
    fn execute<'life0, 'async_trait>(
        &'life0 self,
        args: Value,
    ) -> Pin<Box<dyn Future<Output = Result<ToolResult>> + Send + 'async_trait>>
       where Self: 'async_trait,
             'life0: 'async_trait;

    // Provided methods
    fn output_schema(&self) -> Option<Value> { ... }
    fn param_domains(&self) -> Vec<(&'static str, OptionDomain)> { ... }
    fn invocation_triggers(&self) -> Vec<String> { ... }
    fn spec(&self) -> ToolSpec { ... }
}

Required Methods

fn name(&self) -> &str

Tool name (used in LLM function calling)

fn description(&self) -> &str

Human-readable description

fn parameters_schema(&self) -> Value

JSON schema for parameters

fn execute<'life0, 'async_trait>( &'life0 self, args: Value, ) -> Pin<Box<dyn Future<Output = Result<ToolResult>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,

Execute the tool with given arguments

Provided Methods

fn output_schema(&self) -> Option<Value>

JSON schema describing the structured output this tool attaches to ToolOutput::data, when it declares one. None means the tool emits display text only; authoring surfaces fall back to sample-derived shapes from captured runs.

fn param_domains(&self) -> Vec<(&'static str, OptionDomain)>

Option domains for parameters whose value sets live in runtime config rather than the static schema (channel refs, peer targets, …). Surfaces resolve these through the runtime to render real selectable choices. Default: no domain-typed parameters.

fn invocation_triggers(&self) -> Vec<String>

Lowercase phrases that suggest an inbound message wants this tool — e.g. routing wording or configured destination names for a delivery tool. A pre-turn prefilter may scan the message against these and nudge the model to consider the tool; the tool itself is never called by the prefilter.

Matching contract: case-insensitive, word-boundary — a trigger matches only where it is bounded by non-alphanumeric characters or the message ends (see invocation_trigger_matches). So dev does not fire inside device and ops does not fire inside stops; a trigger only hits as a whole word or phrase. Entries should still be wording that, taken whole, indicates this tool — avoid single common words. A config-derived name that happens to equal a common word can still over-fire; the hint is advisory (the model, not the prefilter, decides to call), so the cost of a false positive is a wasted nudge, never an action. Entries derived from runtime config (aliases, group names) must be computed live per call, never cached across reloads. Default: no triggers; the tool does not participate in prefilter hints.

fn spec(&self) -> ToolSpec

Assemble this tool’s spec. The default recomposes it from parameters_schema() and allocates a fresh Arc per call, so tools with large stored schemas override it to hand out Arc::clone instead (see McpToolWrapper). Delegating wrappers around a dyn Tool MUST forward spec() to the inner tool — relying on this default silently reintroduces a per-iteration deep clone of the inner schema and breaks Arc identity for downstream sharing.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors