THE EDGE AI AND AGENTS ORCHESTRATION PLATFORM.
Author workflows as graphs in a visual builder. Deploy the engine as a Docker image to any Linux edge device. Every decision is visible (inspectable), every run can be replayed in CI (replayable), the action set is finite (auditable), and every AI node can only reach what you wired (bounded) — no cloud dependency.
THE RIGHT INTELLIGENCE AT THE RIGHT LAYER.
Not every task needs a 70-billion-parameter model. ForestHub orchestrates exactly what each task requires.
An industrial workflow is rarely an end-to-end pipeline of the most expensive cloud model. It's a cascade: hard rules where safety matters, classical ML where proven classification is enough, small language models on-device where you need language but the data stays local, and large cloud models only where the extra quality earns the data egress.
On each node, you pick the tier that fits the task. Inference cost drops where it can, data stays where you want it, latency stays in the millisecond-to-second range, and the system no longer depends on a single cloud LLM vendor.
- 01
Rule-based Logic
On-device · Sub-millisecond · Data: stays local
Deterministic if-then logic, thresholds, state machines. Not AI, fully predictable, sub-millisecond. For safety-relevant decisions — "under 80 °C: do nothing, above 90 °C: shut down" — this remains the right tool, because it's auditable without training data.
- 02
Classical Machine Learning
On-device · Milliseconds · Data: stays local
XGBoost, random forests, compact neural nets for anomaly detection and classification. Runs on-device in a few megabytes of memory, millisecond inference. Proven models for the most common industrial tasks — vibration-based machine monitoring, line-side image recognition, predictive maintenance — none of which need a language model.
- 03Sweet spot
Small Language Models (SLMs)
On-device · Seconds · Data: stays local
Models like Phi, Gemma, or Llama in the 1–3B parameter class. Architecturally complete language models, just smaller. They run on-device on industrial PCs or gateways in a few gigabytes of memory, with response times in seconds. Enough for classification, short diagnostic dialogs with the technician, and document RAG — entirely on your hardware.
- 04
Mid-sized Open-Weight Language Models
On-premise GPU · Seconds · Data: stays local
70B-class models such as Llama or Qwen, runnable on NVIDIA Jetson clusters or a dedicated on-premise GPU server. Business value: demanding reasoning on-premise, without data flowing into someone else's cloud. The tier for high-stakes diagnostic work under strict data-protection requirements.
- 05
Frontier Language Models via Cloud
Cloud · Seconds · Data: leaves your network
Claude, GPT, Gemini. Highest quality, best multilingual performance, full provider toolchains. Reserved for the few cases where the extra quality justifies sending data outside — complex customer-facing dialog or multilingual reasoning a 3B model can't carry.
For roughly nine out of ten industrial tasks, a small language model on your own hardware is enough. Frontier models remain the exception, used deliberately — not the default. That is the honest answer to "What does this cost us?" and "Where does our data go?".
VISUAL BUILDER.
Author workflows as graphs on a canvas. Wire sensors, deterministic operations, LLM agents, and actuators into a single artifact — every possible flow is visible at design time.
ENGINE RUNTIME.
The engine is a Go binary in a Docker image — distroless, ~10–15 MB, linux/amd64 + linux/arm64. It loads `workflow.json` and interprets the graph at runtime. No code generation, no recompile per device.
STANDALONE. OFFLINE.
Engine, LLM-proxy, workflow contract, and visual builder ship in github.com/ForestHubAI/edge-agents. Runs without account or cloud. AGPL-3.0 — commercial license on request.
It's the engine behind ForestHub edge agents — bounded by the graph, deployable on your Linux edge devices. Learn about edge agents
YOUR DATA STAYS ON YOUR DEVICE.
No cloud. No compromise.
EVERYTHING YOU NEED TO SHIP.
The visual builder, the runtime engine, and the testing layer — built so engineering teams can author industrial agents, deploy them to Linux edge devices, and prove they behave.
VISUAL BUILDER.
Author workflows as graphs on a canvas. Wire sensors, deterministic operations, LLM agents, and actuators into a single artifact — every possible flow is visible at design time.
Built for technical builders: explicit composition, deny-by-default tool surface, push-based state, branching as a first-class output of LLM nodes. Workflows version like code. Deploys roll back like deployments.
ENGINE RUNTIME.
The engine is a Go binary in a Docker image — distroless, ~10–15 MB, linux/amd64 + linux/arm64. It loads `workflow.json` and interprets the graph at runtime. No code generation, no recompile per device.
Hardware IO (GPIO, ADC, DAC, PWM, Serial, I²C, SPI) goes through Linux kernel APIs. Industrial protocols (MQTT, Modbus TCP/RTU, OPC-UA) are first-party node types. LLM routing covers OpenAI, Anthropic, Gemini, Mistral, plus self-hosted endpoints — and alongside language models, classical ML models (ONNX, TensorFlow Lite) and SLMs are first-class node types, so each step in the same graph can pick the tier that fits the task.
AGENT TESTING & VALIDATION.
Agentic systems must be provably reliable. ForestHub lets you simulate and validate an agent's behavior before it ships to hardware — the critical step from tech demo to engineering product.
Define expected inputs and outputs, test against edge cases, and ensure the agent makes correct decisions under real-world conditions — regardless of whether the AI node behind it is a classical ML model, an SLM, or a frontier LLM. Scenario-based testing keeps agentic logic auditable.
YOUR DATA STAYS ON YOUR DEVICE.
No cloud. No compromise.
ForestHub follows an edge-first architecture. AI inference runs directly on your hardware, which means sensitive data never leaves your device. No cloud uploads, no third-party processing, no data exposure.
For industries with strict data protection requirements - from manufacturing to healthcare - this means GDPR/DSGVO compliance by design. Full control over your data, at all times.
Edge-First Architecture
All AI processing happens on-device. Your data stays where it belongs - on your hardware, under your control.
No Cloud Dependency
The engine runs independently on your edge device. No internet connection required for inference, no recurring cloud costs, no telemetry to a vendor cloud.
GDPR-Compliant by Design
When data never leaves the device, compliance is built in. Ideal for regulated industries and air-gapped environments.
BUILT FOR EU COMPLIANCE.
EU AI Act
Risk classification, documentation, and human oversight built in from day one. Full traceability for high-risk AI system obligations.
Cyber Resilience Act
Secure-by-design updates, SBOM support, and vulnerability management. CRA-ready for connected products out of the box.
GDPR / DSGVO
All processing on-device. No personal or operational data leaves your hardware. GDPR compliance by design — not by policy.
RED & CE Marking
Edge deployment on certified hardware ensures compliance with the Radio Equipment Directive and CE marking requirements for connected devices.
BUILD ONCE, DEPLOY EVERYWHERE.
The workflow is hardware-agnostic. The engine targets `linux/amd64` and `linux/arm64` — gateways, NUCs, NVIDIA Jetsons, Raspberry Pis, industrial PCs all run the same Docker image. Migrating from one hardware generation to the next means redeploying the same image, not rewriting.
Linux-Edge Native
Distroless Docker image, `linux/amd64` + `linux/arm64`, ~10–15 MB. Runs on any Linux edge device — Yocto, Buildroot, OpenWrt, Balena, standard distributions.
Workflow as Artifact
Same graph, gateway or NUC or Jetson. No build matrix, no per-device variants. Deploy the engine, load the workflow.
Industrial Protocols First-Party
MQTT (Mosquitto + auth), Modbus TCP/RTU, OPC-UA — node types in the graph, not bolted-on plugins. Triggers, publishers, and subscribers wire into the same canvas.
WHAT YOU CAN BUILD WITH FORESTHUB.
From industrial diagnostics to environmental sensing - ForestHub enables Edge AI for a wide range of applications.
Intelligent Service Technician
AI-powered on-device diagnostics for industrial equipment. Real-time fault detection and guided troubleshooting without cloud dependency.
Building Automation
Local AI for HVAC, lighting, and energy management. Process sensor data on-device for real-time building control.
Predictive Maintenance
Detect anomalies and predict failures before they happen. Vibration and temperature analysis running directly on edge hardware.
Autonomous Sensing
Intelligent data filtering at the sensor node. Drastically reduced data transmission through on-device classification and event detection.
And Much More...
Quality inspection, fleet monitoring, environmental sensing, wearable AI - the platform adapts to your specific Edge AI challenge.
FREQUENTLY ASKED QUESTIONS.
The ForestHub Orchestration Platform launched on March 13, 2026. Register for free at app.foresthub.ai — no credit card required.
The Extended Survey helps us understand your use case, hardware requirements, and experience level. Completing it helps us tailor the platform to real-world needs and prioritize the features that matter most to you.
ForestHub is free to get started. No credit card required, no hidden costs. We offer tiered pricing plans for teams and advanced features — check app.foresthub.ai for current details.
The engine is a Go binary in a distroless Docker image, ~10–15 MB, targeting `linux/amd64` and `linux/arm64`. It runs on any Linux edge device — gateways, NUCs, NVIDIA Jetsons, Raspberry Pis, industrial PCs. No vendor-specific build, no cross-toolchain.
LangChain and OpenAI Agents SDK make the LLM's loop the program — give the LLM tools and let it iterate until done. ForestHub inverts that: the graph is the program, the AI node is one among many. And a node need not be a frontier LLM: in the same graph ForestHub orchestrates rule-based logic, classical ML, on-device small language models, and large cloud models — pick the tier that fits each node. The loop-first pattern is still available as a single agent node inside the graph — bounding is opt-in, so we contain the loop-first model rather than replace it.
Yes. The engine, LLM-proxy, language-neutral workflow contract, and the visual builder all live in github.com/ForestHubAI/edge-agents. The repository runs standalone and offline — no external services or accounts required. Dual-licensed: AGPL-3.0-only for the public release, commercial license on request for use cases that need a non-AGPL track.
READY TO BUILD THE FUTURE?
The ForestHub Platform is live. Register now and deploy your first workflow to a Linux edge device.
$ git clone https://github.com/ForestHubAI/edge-agents