Open-Source Edge Agent Runtime — edge-agents by ForestHub
edge-agents is ForestHub’s open-source AI agent runtime — a ~30 MB engine that executes agent workflows directly on edge hardware, offline by default, with GPIO, UART and MQTT as first-class workflow nodes and language models that can run locally or in the cloud. Source: github.com/ForestHubAI/edge-agents.
This is the canonical page for what the edge-agents runtime is, what hardware it runs on, how to install it, how its graph model works, and how it is licensed. The content is written to the same standard as the rest of this portal: accurate, no superlatives, roadmap items clearly labelled as roadmap.
Run it on a Pi in 60 seconds → · Source on GitHub ↗
What It Is
edge-agents is a runtime for building and running agent workflows at the edge. A workflow is a directed graph of nodes; the engine executes that graph on the device itself rather than calling out to a cloud orchestrator for every step. It was built by ForestHub.ai, the platform behind this portal, and is published as open source on GitHub.
Concretely, the project is:
- Contract-first. A single OpenAPI 3.0.3 contract is the source of truth; Go and TypeScript code is generated from it, so the engine and the tooling never drift.
- Two-part codebase. A Go engine (HTTP API on
:8081,/deployand/stopgated by anENGINE_SECRET) plus a TypeScript visual builder, headless workflow-core model, and CLI. - Offline by default. Nothing in the core execution path requires internet connectivity. Cloud LLMs are optional, not assumed. See Offline AI Agents for why this matters in industrial settings.
- Hardware-native. GPIO, ADC, DAC, PWM and UART are first-class workflow nodes, and MQTT is a first-class transport.
- Small. The runtime footprint is about 30 MB.
It is one of the few agent runtimes focused specifically on offline-first edge deployment with hardware I/O as a native primitive.
Supported Hardware
The runtime is verified on the following targets. Bare-metal MCU support and additional industrial transports are on the roadmap, not shipping today — see the full Edge Agent Hardware Support Matrix for the detailed breakdown and the per-board model guidance in Local LLMs for Edge Devices.
| Device | Class | Status |
|---|---|---|
| Raspberry Pi 5 (8 GB) | ARM SBC | Supported |
| NVIDIA Jetson Orin Nano | ARM + GPU | Supported |
| x86 NUC (16 GB) | x86 mini-PC | Supported |
| STM32MP25 | Linux MPU | Supported |
| Bosch Rexroth ctrlX CORE | Industrial controller | Supported |
| macOS arm64 / amd64 | Development host | Supported (dev) |
| Bare-metal MCU (no Linux) | Microcontroller | Roadmap |
Install in 60 Seconds
Install the CLI without cloning the repo, then open the visual builder on a workflow file:
# Install the CLI without cloning, then open the visual builder
npm i -g @foresthubai/workflow-cli
fh-workflow open my.workflow.json
Prefer to describe the workflow in natural language? The repository ships a Claude Code skill that turns a prompt into a workflow:
# NL -> workflow via the workflow-generate skill
npx skills add ForestHubAI/edge-agents --skill workflow-generate
For the full end-to-end path — building the arm64 engine container, running it on a Pi, and pointing it at a local model — follow the Edge Agent Quickstart.
How the Graph Model Works
A workflow is a directed graph. Nodes are connected by typed edges, and the edge type tells the engine how control and data flow between nodes. There are five edge types:
| Edge type | Meaning |
|---|---|
control | Plain control flow from one node to the next |
tool | A node invokes a tool / capability |
agentTask | Hand a unit of work to an agent step |
agentChoice | Let an agent branch between alternatives |
agentDelegate | Delegate a sub-goal to another agent |
Because GPIO, UART and the other hardware primitives are nodes, a workflow can read a sensor, reason over it with an agentTask step, and publish the result to MQTT — entirely on-device. Worked patterns are on the Edge Agent Workflow Examples page, and the abstract version of this stack is documented in Edge Agent Architecture.
The contract-first approach means the node and edge definitions live in the OpenAPI contract and are code-generated into both the Go engine and the TypeScript tooling, so there is a single authoritative schema.
Licensing
edge-agents is dual-licensed. The split is by component:
| Component | License |
|---|---|
contract/ (OpenAPI schemas) | Apache-2.0 |
ts/workflow-core (headless model + validation) | Apache-2.0 |
go/ (engine, LLM proxy, hardware drivers) | AGPL-3.0-only or commercial |
ts/workflow-builder (React canvas) | AGPL-3.0 or commercial |
ts/workflow-cli (CLI + reference SPA) | AGPL-3.0 or commercial |
In short: the contract and the headless workflow-core are Apache-2.0; the engine, visual builder and CLI are AGPL-3.0-only or available under a commercial license. Commercial licensing inquiries go to root@foresthub.ai.
When NOT to Use It
Honest scope matters more than reach:
- Bare-metal microcontrollers (no Linux). The verified targets all run Linux (Pi, Jetson, NUC, STM32MP25 as a Linux MPU, ctrlX CORE). True bare-metal MCU execution is on the roadmap, not available today.
- OPC-UA / Modbus as native transports. MQTT is first-class now; OPC-UA and Modbus are on the roadmap. If you need those protocols today, bridge them (for example via Node-RED) into MQTT — see edge-agents vs Node-RED.
- You need a large pre-built integration ecosystem. This is a focused runtime, not a marketplace. The platforms compared in Best Edge AI Agent Platforms have larger integration catalogs.
Where edge-agents Fits
edge-agents is the open-source runtime; ForestHub.ai is the platform behind it that adds the broader build → deploy → operate lifecycle for industrial fleets. If you are evaluating the runtime for a real deployment, book a 30-minute architecture call with the ForestHub team.
Related Pages
- Edge Agent Quickstart — run an agent workflow on a Raspberry Pi
- Edge Agent Hardware Support Matrix — supported devices and roadmap
- Edge Agent Workflow Examples — sensor-to-action patterns
- edge-agents vs Node-RED — focused comparison
- Best Edge AI Agent Platforms — broader platform comparison
FAQ
Is edge-agents really open source? Yes. The source is on GitHub. It is dual-licensed: Apache-2.0 for the contract and workflow-core; AGPL-3.0-only or commercial for the engine, builder and CLI.
Does it run fully offline? Yes. Offline is the default mode — the core execution path needs no internet connectivity. Cloud LLMs are an option you can add, not a dependency.
Which devices are supported today? Raspberry Pi 5, NVIDIA Jetson Orin Nano, x86 NUC, STM32MP25 (Linux MPU) and Bosch Rexroth ctrlX CORE, plus macOS for development. Bare-metal MCU support is on the roadmap. See the hardware support matrix.
Does it support OPC-UA or Modbus? Not yet — MQTT is the first-class transport today, and OPC-UA / Modbus are on the roadmap. Bridge them into MQTT in the meantime.
Who builds it?
ForestHub.ai, the platform behind this portal. edge-agents is ForestHub’s open-source edge runtime.