What problem does it solve?
Building LLM applications that react to webhooks, timers, queues, or task completions requires careful handling of event ingestion, identity trust, state persistence, and safe model invocation. This Skill provides the patterns and rules for wiring event sources through an inbox and routing layer into AxGen, AxAgent, and AxFlow programs without letting untrusted event data directly trigger model calls.
Core Features & Use Cases
- Event Routing with Explicit Actions: Routes declare
observe, invalidate, wake, or resume so only deliberate actions invoke a model, with dead-lettering of invalid inputs before invocation.
- State Persistence and Continuations: Persist run state and outputs, register continuation correlation keys, and resume owning programs when task events complete, fail, or require input.
- Adapters for MCP and UCP: Integrate MCP catalog subscriptions and UCP webhook sources with verified identity, signature checking, and replay protection.
- Use Case: Imagine an incident-response agent that must wake when an
incident.created webhook arrives, persist its triage output before delivering it to a sink, and resume when a human task completes. This Skill defines the exact source, route, target, store, and sink configuration to do that safely.
Quick Start
Set up an Ax event runtime that wakes my triage agent whenever an incident.created webhook event arrives and persists the result to a sink.