Embedded Agent Hardware Support: What Runs an AI Agent Runtime (Pi, Jetson, STM32MP25, ctrlX CORE)
Embedded Agent Hardware Support
An embedded agent runtime can run on a range of hardware classes, but the practical dividing line is whether the target runs a Linux-class OS or is a bare-metal microcontroller. Today, containerized agent runtimes target Linux-class embedded and industrial edge hardware; bare-metal MCU deployment is a separate, harder problem that most runtimes — including the open-source edge-agents runtime — treat as roadmap, not current support.
This page is a hardware-selection reference. It first describes the hardware classes and their requirements, then gives the verified target list for a concrete runtime so the distinction is unambiguous.
Hardware classes for an embedded agent runtime
| Hardware class | Typical RAM | Architecture | OS | Agent-runtime fit |
|---|---|---|---|---|
| Industrial PC / edge server | 8–32 GB | x86_64 | Linux | Full — containerized runtimes run natively |
| Single-board computer (SBC) | 4–8 GB | ARM64 | Linux | Full — e.g. Raspberry Pi 5, Jetson Orin Nano |
| Linux MPU module | 0.5–2 GB | ARM Cortex-A | Linux | Supported — e.g. STM32MP25 (Cortex-A35) |
| Industrial controller (Linux) | 1–4 GB | ARM64 | Linux (vendor) | Supported — e.g. Bosch Rexroth ctrlX CORE |
| Bare-metal microcontroller | 64 KB–2 MB | ARM Cortex-M | RTOS / bare-metal | Not yet — see roadmap note below |
The split is about the operating system, not raw compute. A Linux-class target can host a containerized runtime, a process scheduler, a full network stack, and a model proxy. A Cortex-M microcontroller running an RTOS or bare-metal superloop cannot host a container and has a different, lighter deployment model — see the embedded agent architecture runtime table for the RTOS-task and bare-metal-loop patterns that apply there.
Verified targets: the edge-agents runtime
The open-source edge-agents runtime publishes a list of hardware it has been verified on. It is a ~30 MB containerized runtime, which is why every verified target is Linux-class:
| Verified target | Class | Notes |
|---|---|---|
| Raspberry Pi 5 | SBC (ARM64) | Common prototyping and edge target |
| NVIDIA Jetson Orin Nano | SBC (ARM64) | Adds GPU for local inference |
| STM32MP25 | Linux MPU (Cortex-A35) | Linux MPU, not a bare-metal MCU |
| Bosch Rexroth ctrlX CORE | Industrial controller | Linux-based industrial control platform |
| x86 NUC | Industrial PC (x86_64) | Edge-server-class deployment |
Supported today vs roadmap
- Linux-class embedded and industrial edge (supported today): the five targets above. The runtime executes locally on each — no cloud control plane required.
- Bare-metal microcontroller / Cortex-M (not supported, on the roadmap): this runtime does not deploy to bare-metal MCUs. Note that STM32MP25 is an MPU running Linux, not a Cortex-M microcontroller — it belongs in the supported column. For the MCU-specific developer view, see the sister hub mcuagent.com.
This honest split is the point of the page: an engineer choosing hardware should know exactly where a containerized agent runtime fits and where it does not.
Source and next steps
- Repository: github.com/ForestHubAI/edge-agents — verified hardware list and install instructions.
- Platform: ForestHub.ai — the platform behind the runtime.
- Deploy it: Embedded Agent Runtime Quickstart.
Related pages
- Open-Source Embedded Agent Runtime — what the runtime is.
- Embedded Agent Runtime Quickstart — deploy onto these targets.
- Embedded Agent Architecture — runtime models per hardware class.
- Industrial Embedded Agents — industrial-controller context.
This reference is part of a ForestHub.ai knowledge hub. Book a meeting →