Last week, Anthropic announced a new specification built on top of the foundations and learnings of MCP, Model Hardware Specification. From their announcement post, Model Hardware Standard (MHS) is a shared specification for AI agents to safely operate physical devices. Initially previewed for physical lab and manufacturing hardware (robot arms, qPCR machines, microscopes), this specification may open the door for AI to dramatically transform and reshape our physical world in unexpected ways.
Though MHS is similar to MCP, the key difference is that this specification is an entryway for manipulating real-life equipment and robotics, not just software. As a result, the potential security risks and surface area are very different.
We think this is where the edge is heading in a post-agentic world: not a layer sitting on top of software, but shared infrastructure for software and hardware systems together.
When we built and shipped our edge MCP prototype recently, we were building on an architectural foundation of exceptional process execution security, network resilience, and agent performance. Today, we’ve taken the first steps to extend those concepts of securing and scaling agentic operations into the world of hardware. You can see our research prototype here, which we’ll keep up to date as more of the specification and standard materializes.
Performance as Safety #
Let's zoom out and discuss the difference between securing exclusively agentic software systems and agentic software/hardware hybrid systems.
In an agentic commerce experience, a slow or dropped connection between an agent and a checkout API means someone clicks "buy" twice. This is a terrible customer experience, and one that causes customer and operator pain.
A delayed connection between an agent and a robot arm or life sciences experiment means an agent that isn't sure whether its command landed. "Retry the move" is a materially worse failure mode than retrying the form submission." In robotic manufacturing, a retry might cause the agent to break the physical material the robot is manipulating. In a lab, the risks are even more extreme: actions like a qPCR run have high stakes for advancing global health research, and a mishandled request can ruin the experiment.
Exploring a Performant and Secure implementation #
Since MHS is so adjacent to MCP, we forked fastly/edge-mcp, our recently published implementation of a stateless MCP server for Compute, and started extending the system to explore MHS capabilities.
This was an MHS implementation tailwind: edge-mcp's protocol engine and Fastly Compute bindings needed no changes. Auth, scope checks, and schema validation already run before a tool handler executes, so all the new MHS-specific security and integration prototyping are plugged into the existing handler interface. Specifically, we built:
Safety limit enforcement: every tool call is checked against the target device's declared safety limits before it's forwarded. A temperature bound, a set of allowed axis values, whatever the device metadata says.
Per device quotas: In addition to traditional edge rate limits, an authorized agent can be throttled, because repeated commands can wear out hardware even when every individual command is legitimate.
Audit trails: Every decision gets logged with a hashed identity and a correlation id, with the goal of being able to reconstruct what happened without storing raw tokens or arguments.
Monitoring the Standard as it Develops #
At the time of writing, the model hardware standard is still very much in an early stage and in development. What we've built is a prototype proposal for what would be ideal for a secure edge implementation.
But make no mistake, we believe this is the direction of the agentic edge: emerging software, network, and hardware controls in a way that demands exceptional end-to-end security and performance. We believe that we are uniquely positioned to be the best platform for this layer within the modern agentic architecture, especially as the line between software and hardware begins to blur.
If you would like to collaborate with us on this work, or if you would like us to work with your technical teams and leadership to take advantage of the agentic transformation moment that we're in, reach out.