Anthropic's Model Hardware Standard lets AI agents run lab robots
An open 'MCP for machines' preview lifted a quantum lab's laser relock from 58% to 99.3% success — but it's gated. Plus htmx 4.0 and repo-free Cursor agents.
Copy markdown
It's MCP, but for physical machines
MHS gives agents a standard driver with read/write primitives — get a temperature, set a laser's power — so they can discover and drive any programmable instrument over MCP, a CLI, or plain code, with no per-device glue. Vendors already onboard include Universal Robots, Tecan, Doosan, and AWS Strands Robots.
The lab numbers are real
QuEra's quantum-computer laser relock went from 58% success in ~150s to 99.3% in under 15s. Carnegie Mellon cut dose-response runs ~3x and wired up hardware in 8 hours instead of weeks; Genentech collapsed integrations from months to minutes.
The catch: gated, and no SDK yet
MHS isn't open-source and ships no public repo or SDK — you apply for the research preview at modelhardwarestandard.com. Anthropic also warns Claude lacks 'physical intuition' (it needed coaching on bubble formation in a protein assay), and formal safety evals are still underway.
htmx 4.0 'The Fetchening' lands
The hypermedia library's first major bump in years rebuilds on the native fetch() API and adds DOM-morphing swaps, with breaking migration changes. If htmx powers your AI app's frontend, budget time before you upgrade.
Cursor cloud agents drop the repo requirement
You can now hit 'Start from scratch' and prompt a Cursor cloud agent with no GitHub repo — it spins up an Origin repo, gives a browser live-preview with design mode, and can publish straight to a Vercel URL. Handy for zero-to-prototype weekend builds.