@agentium/edge
Edge deployment for Agentium — IoT toolkits (GPIO, sensors, camera, BLE, servo, system) plus an edge runtime with resource monitoring, presets, and Ollama helpers. Designed for Raspberry Pi but works on any Linux SBC.
Install
npm install @agentium/edge
Hardware toolkits load their native bindings on demand. Install only what you use, e.g.:
npm install onoff # GPIO
npm install i2c-bus # sensors
npm install @abandonware/noble # BLE
Quick Start — Ollama-backed local agent on a Pi
import { Agent } from "@agentium/core";
import { EdgeRuntime, edgePreset, ollama } from "@agentium/edge";
const runtime = new EdgeRuntime(edgePreset("pi-5-balanced"));
await runtime.start();
const agent = new Agent({
name: "edge-bot",
model: ollama(runtime.recommendedModel),
instructions: "You are an edge AI assistant running on-device.",
});
console.log(await agent.run("hello"));
IoT Toolkits
Each toolkit exposes typed tools (with Zod schemas) the agent can call:
| Toolkit | Example tools |
|---|---|
GpioToolkit | gpio.write, gpio.read, gpio.toggle |
SensorToolkit | sensor.read_temperature, sensor.read_humidity (I²C / 1-Wire) |
CameraToolkit | camera.snapshot, camera.start_stream |
BleToolkit | ble.scan, ble.connect, ble.read_characteristic |
ServoToolkit | servo.move_to, servo.center |
SystemToolkit | system.cpu_temp, system.uptime, system.shutdown |
import { Agent } from "@agentium/core";
import { GpioToolkit, SensorToolkit } from "@agentium/edge";
const agent = new Agent({
name: "garden-bot",
model: ollama("llama3.2:3b"),
toolkits: [
new GpioToolkit({ pumpPin: 17 }),
new SensorToolkit({ moistureBus: 1 }),
],
});
await agent.run("Water the plants if soil moisture is below 30%.");
Edge Runtime
EdgeRuntime continuously monitors CPU, RAM, temperature, and storage so you can react to resource pressure (downsize the model, defer non-critical tools, sync to cloud, etc.).
const runtime = new EdgeRuntime(edgePreset("pi-zero-2-low-power"));
runtime.on("threshold:cpu", (s) => console.warn("CPU pressure", s));
await runtime.start();
Cloud Sync
EdgeCloudSync ships local agent runs / sensor logs to a central endpoint with offline queueing — useful for fleets of Pis with intermittent connectivity.
Documentation
Full docs at docs.agentium.in
Community
Join the conversation on Discord.
License
MIT
Sync and camera behavior
Concurrent flush calls share the active delivery batch. Events queued during that request remain persisted for the next batch; only acknowledged IDs are removed. Failed deliveries retain their batch for retry, so receivers must handle duplicate delivery. Camera capture returns the complete base64 image before deleting its temporary file; applications still choose how to present the image to a multimodal model.