Engine at a glance
Engine is an experimental, local-first runtime for durable goals across software and physical worlds. It combines a durable Heart, exactly one active and replaceable general executive brain, zero or more specialist brains, and installable world plugins.
A request can be immediate—“turn this light off”—or become a durable intent: ACHIEVE an outcome once, or MAINTAIN it as the world changes. Engine persists that intent independently of the conversation, model, or process that first expressed it.
interaction -> intelligence -> Engine -> target / world
chat or API model/planner durable software, home,
runtime business, energy,
or machine
The intelligence layer is replaceable cognition, not operational authority. The design starts with a current world and a desired world that continue to exist without a model context:
durable current world + durable desired goal
|
v
event or poll wakes the Heart
|
v
observe again and assess
|
v
brain/specialist proposes
|
v
schema -> policy -> authorization
|
v
execute -> observe again
|
v
oracle -> receipt -> experience
A brain may choose and propose. The Heart maintains the cycle, state, and causal history. Policy grants or denies authority. A target-specific executor acts. An event is only a wake-up hint, and an API acknowledgement is only an execution fact. Only a fresh observation and an effect oracle may establish what actually happened within their current coverage.
The essence in four statements
- Heart means continuity. Goals, snapshots, receipts, and experience remain durable outside the model session.
- Brain means replaceable cognition, not authority. A deterministic planner or model-backed executive chooses the next cognitive step; specialists provide bounded advice.
- Plugins represent worlds with their own semantics. Providers observe; controllers translate; executors act; oracles measure effects.
- Success requires independent evidence. Model output or an API acknowledgement is never enough.
What to expect today
| Status | Meaning in this documentation |
|---|---|
| Implemented | Present in the current action lifecycle and v3 plugin contracts |
| Fake/simulation-tested | Proved automatically, but not evidence of physical safety |
| Live read-only | Observed against a real target without mutation |
| Roadmap | A direction or hypothesis, not a feature available today |
The current repository contains:
engine-heart, with the Heart, durable world store, policy, learning, and routines;engine-sdk, withengine.plugin/v3, v2 compatibility, scaffolding, and conformance tooling;engine-runtime, with the composition root, discovery, lease, and theengineCLI;- a warehouse reference plugin, context plugin, Homey plugin, and opt-in bounded ntfy lifecycle observer;
- tests for reconstruction, stale/denied/malformed cases, immediate and task lifecycles, and bounded learning.
The current evidence boundary is deliberately narrower than the vision: Homey observation is Live read-only, while its v2 mutation path is Fake/simulation-tested. STREAM, multiple general executive brains, trained mini-brains, and physical-safety certification have not been demonstrated. The end goal works through five possible applications and labels each one at this boundary.
Choose a reading path
I want to understand Engine
I want to understand every capability and boundary
I want to build
- Quickstart
- Plugin interface v3
- SDK reference
- CLI reference
- Multiple plugins and brains
- Plugin checklist
I want to position Engine
The fixed checksum
LLM proposal != authority
prediction != observation
policy != physical safety
deliberation != realtime control
simulation evidence != real-world certification
generic lifecycle != generic device semantics
state != weights
imagine != execute
These statements are not slogans added after the fact. They define which code, tests, and claims belong in Engine.