Open protocol · draft 0.2

What is the Open Model Hardware Protocol?

OpenMHP (MHP) is an open standard for connecting AI agents to physical devices: microscopes, liquid handlers, robot arms, furnaces, lasers. It gives an agent one way to find an instrument, learn how to use it, operate it safely, and hand it long-running work.

Wire format JSON-RPC 2.0Transports stdio · HTTPWorks with any MCP-capable agent harnessReference implementation Python

Think of MHP the way you think of MCP. MCP gives an AI application a standard port for data and software tools. MHP gives it a standard port for hardware. An instrument that speaks MHP can be operated by any agent harness, and a harness that speaks MHP can operate any instrument, without a bespoke integration in between.

The harness on the left is whatever you already run. MHP sits in the middle. The hardware on the right keeps whatever control layer it has.

What can MHP enable?

An agent in any harness finds an idle thermocycler in bay 12 by describing what it needs, reads the owner's operating notes, and runs a 30-cycle PCR without a human writing a line of integration code.
A drug-discovery workcell runs overnight: the agent hands a script to the devices, the liquid handler, plate mover and reader carry it out, and only the summary comes back to the model.
A quantum-computing lab lets an agent sweep a laser calibration, with the limits the physicist wrote in the device package enforced on every setpoint.
A factory cell exposes a furnace and a robot arm through the PLC they already have; the agent may set a temperature within range, may never open the door without a human, and can always stop everything.

Why does MHP matter?

Instrument owners write one device package and every agent harness can use the instrument. The safety limits they write are enforced by the driver, whatever the agent asks for.
Harness and agent developers get one integration for all hardware: eight MCP tools that stay the same from two devices to two thousand, with the agent's context cost held flat.
Scientists and operators get agents that can run real experiments, with a human confirmation step wherever the owner said one belongs, and an emergency stop the agent is always allowed to press.

Works with any agent harness

MHP is not tied to one model or one product. The reference bridge, mhp-mcp, is an ordinary MCP server, so any MCP-capable harness can operate an MHP lab: Claude Code, Codex, OpenClaw, Hermes, Claude Science, Open Science, and custom agents built on any SDK. Harnesses that would rather skip MCP can use the MHP client library or the CLI directly. Three Agent Skills ship with the protocol so that skills-capable harnesses know how to onboard hardware and operate it safely without bespoke prompting.

Start building

QuickstartRun two simulated instruments and connect your harness in five minutes. Add an instrumentWrite a device package so any agent can control your hardware safely. AdaptersBring SiLA 2, PyLabRobot, MADSci, OPC UA or ROS 2 fleets under MHP in a few lines each. SpecificationPrimitives, safety model, directory, transports, conformance.

Where OpenMHP came from

Getting several devices in a lab or on a factory floor to work together has always been hard, and putting an AI agent in charge of them is harder still. Each instrument has its own programming interface. There has been no standard way to integrate them, and once they are wired up there is no common way for them to share data with an agent, or to let the agent operate them safely. The knowledge that makes a device safe to use, such as how heavy an arm is or when a lid is too hot to open, sits in paper manuals, on someone's laptop, or in the head of the person who runs it.

Anthropic solved the equivalent problem for software when it open-sourced the Model Context Protocol. MCP gave any AI application one standard way to reach digital tools, and the ecosystem followed. Agent Skills, also open-sourced, showed how to package procedural knowledge so an agent loads only what a task needs. OpenMHP is the same idea carried across to hardware: a shared, open protocol so that any AI agent can connect to lab instruments and machines the way MCP connected it to digital tools, and run experiments agentically, from routine drug-discovery assays to laser calibration on a quantum computer.

Three choices follow from that origin.

A standard driver with a small set of primitives. Read a signal, write a setting, invoke an action, stop. Any device can implement them, and each device becomes discoverable in a standard format, so devices and agents find each other across a network without a bespoke translator in between.
The device package holds what the manual never did. Owners write the characteristics an agent cannot infer from code, in plain language, either by hand or by talking to an agent that interviews them. From that the driver serves a card, operating instructions and an enforced safety envelope: what can be measured, what can be adjusted, and what limits will be applied whatever the agent asks for.
Three ways to drive it, working together. MCP for exploring and deciding; a command line for checking and scripting; code files that chain driver commands across devices when work must run for hours or faster than an agent can reason. The agent sequences steps across instruments, watches results, adjusts as conditions change, and hands the devices a script when they should carry on without it.

MHP also takes two lessons from MCP's first year. Listing every capability up front bloats the agent's context and makes selection worse, so devices load progressively: a search returns cards, choosing a device loads its instructions, and detailed specs and reference files load only when asked. And examples teach conventions that schemas cannot, so every action and every bridge tool carries them. Both are described in the Scale section of the specification.