Anthropic MHS Lets Claude and AI Agents Control Lab Hardware
Anthropic's Model Hardware Standard links Claude and other AI agents to lab devices, promising faster setup, natural language control and safety limits.
Summary
Anthropic unveiled the Model Hardware Standard, or MHS, research preview on August 27, 2026, creating standardized drivers and a common data format through which AI agents or conventional software can control disparate networked devices without custom translators. Anthropic says early scientific-partner testing over the past year cut integrations that normally take weeks or months to hours or minutes. Technical staffer Alek Kemeny conceived the approach after neuroscientist Arco Bast coordinated rotating lasers, microscopes, cameras and other equipment for memory-formation research at HHMI Janelia Research Campus in Ashburn, Virginia.
Command-line prompts and API files provide direct real-time control, while connection through Model Context Protocol adds natural-language operation, stepwise reasoning, live parameter changes and, in some cases, autonomous recovery from hardware errors. Anthropic demonstrated Claude calibrating a laser from camera feedback, focusing and repositioning a microscope after analysis, and reasoning without task-specific training how a robotic arm could pick up an aluminum can. Models can also write and adapt API scripts that sequence instruments. Reference tags describe device weight, range, adjustable parameters, measurements and enforced safety limits. Preview partners Amazon Web Services' Strands Robots, Hugging Face's LeRobot, Raspberry Pi, Automata and Universal Robots will help build safety evaluations and best practices before Anthropic turns MHS into an open source, agent-agnostic standard.
Positives
- Early scientific-partner testing reduced device integration from weeks or months to hours or minutes, Anthropic says.
- Model Context Protocol lets scientists control equipment through natural language while models adjust experimental parameters in real time.
- Claude calibrated lasers, repositioned microscopes and guided a robotic arm without task-specific training for each demonstrated procedure.
- Standardized tags communicate device weight, range, parameters, measurements and enforced safety limits to unfamiliar AI models.
- Amazon Web Services, Hugging Face, Raspberry Pi, Automata and Universal Robots are helping prepare MHS for broader adoption.
Risks & concerns
- MHS remains a research preview, with open source and agent-agnostic availability planned only after further development.
- Autonomous recovery from hardware failures works only in some cases, limiting fully unattended operation.
- Safety evaluations and best practices for AI-controlled physical equipment are still being developed with preview partners.
- Models may lack prior experience with connected devices, making accurate constraint tags and reference files essential for safe operation.