Spectre is an OTP-native runtime for agents whose routing, state, policies, durability, and side effects remain explicit. This page is the shortest path from a new Mix project to a correct mental model of the runtime.
Your first path
Read these four pages in order:
- Overview explains what Spectre is, when it fits, and the safety boundary it preserves.
- Installation adds the stable package, formatter metadata, and read-only diagnostics.
- Getting Started builds an Agent and follows one turn through routing, policy approval, and explicit host execution.
- Architecture establishes ownership, trust boundaries, Runs, Instances, and the separation between conversational and operational work.
Do not start with every extension or durable subsystem. A deterministic local turn is enough to learn the public boundary; add persistence and integrations only when the application needs them.
Continue by goal
| Goal | Continue with |
|---|---|
| See complete applications | Two Realistic Agents |
| Choose routes and stage side effects | Routing and Actions |
| Own durable subject state | Agent Instances and Subjects, then Resumable Runs |
| Stream an Instance-owned model response | Streaming inference |
| Persist nondeterministic boundary evidence | Boundary receipts |
| Run long or recurring procedures | Work, Vigil, and the Operational Runtime |
| Deploy and verify a release | Production Operations and Testing |
| Integrate another package or host runtime | Integration Boundaries and Stack |
| Look up functions and compatibility | Public API and the Public API Manifest |
| Upgrade an existing application | Read the current release notes, then the Run v3 and Instance v3 guides when applicable |
The guide for LLMs and coding agents is a separate, machine-oriented contract. Human readers should follow the path above first.