Claude Comes to CarPlay as Anthropic Tests AI Control of Lab Hardware

Claude’s expansion beyond chat is taking two very different forms. In Apple CarPlay, it is a voice-first conversational app with strict platform limits. In laboratories and advanced manufacturing, Anthropic’s Model Hardware Standard (MHS) is an early research preview for letting AI agents work with programmable physical instruments under defined controls. 171

What drivers can do with Claude in CarPlay

Claude is available through the Claude iPhone app in Apple CarPlay. A driver can open the app from the CarPlay interface and ask spoken questions through the vehicle’s display and voice controls, receiving voice-based answers without handling the phone. Recent conversations may also be available for continuation. 1718

The practical use case is conversation: asking for an explanation, thinking through an idea, or continuing a prior Claude exchange while keeping interaction voice-led. It is not a full dashboard version of Claude’s visual chat interface.

The important CarPlay restrictions

Apple’s CarPlay support for conversational AI is deliberately narrow. Third-party chatbot apps must use Apple’s designated voice-control screen and obtain the relevant CarPlay entitlement. 1723

More importantly, Claude cannot control vehicle systems or iPhone functions. It is not a replacement for controls such as navigation, climate, music, calls, messages, or phone settings. Users also need to launch Claude from the CarPlay interface; it does not have a dedicated in-car wake-word invocation. 1719

Those limits make Claude a hands-free question-and-answer companion, not a vehicle assistant with privileged access to the car.

How Claude compares with other CarPlay chatbots and Siri

Claude is one of several third-party chatbot choices in CarPlay. Reporting identified it as the fifth major chatbot app on the platform, after ChatGPT, Perplexity, Grok, and Meta AI. 18

That distinction matters: these are optional conversational apps operating within the same CarPlay constraints, rather than assistants integrated into the operating system’s vehicle controls. Apple’s own Siri AI is expected to reach CarPlay with iOS 27, giving Apple a platform-native AI offering alongside third-party options. 22

What the Model Hardware Standard is for

MHS is a shared specification for AI agents to operate physical devices. Anthropic is opening it as a research preview to an initial group of scientific research labs and advanced manufacturers. The company says the standard can support work across instruments including microscopes, liquid handlers, robotic arms, and quantum-computing equipment. 1

The goal is interoperability. Rather than requiring an agent to learn a separate vendor-specific interface for every machine, MHS describes a common driver layer built around standardized operations such as reading a device state and writing an approved setting. A temperature sensor, for example, could expose a readable value, while a controller could accept a bounded temperature setpoint.

Device discovery and operational context

A hardware description can give an agent structured information about what an instrument is, what it can measure or change, and what constraints govern its use. The approach also supports natural-language descriptions of capabilities, adjustable parameters, physical properties, and operating limits.

That context is intended to help an agent discover devices on a network and reason about their permitted operations without treating every machine as an unfamiliar, bespoke API. Agent harnesses can connect through standard interfaces including the Model Context Protocol (MCP), command-line tools, or APIs.

From live reasoning to repeatable scripts

MHS is designed not only for one-off actions but for orchestrating multi-instrument workflows: monitoring live readings, adjusting parameters, and coordinating equipment as results change. Once a workflow is validated, it can be turned into a deterministic device-command script rather than relying on open-ended model reasoning for every repeated low-level action. 1

That is a meaningful safety and reliability distinction. A model may help discover or refine a procedure, while the repetitive execution path can be made more predictable.

The laser-alignment demonstration

Anthropic reports a demonstration in which Claude adjusted a laser, inspected the result through a camera, iterated to learn how adjustments affected beam position, and then wrote a deterministic script that could align the laser with a single command. 1

Anthropic also says Genentech used MHS in a proof of concept to automate a BCA protein assay, a workflow that requires coordinated lab instrumentation. 1

How Anthropic is managing physical-world risk

Anthropic is not presenting MHS as an unrestricted autonomy product. Its rollout is limited to a research-preview group of labs and advanced manufacturers while the company develops safety evaluations and operating practices ahead of a planned open-source release. 1

The framework’s safeguards are meant to be layered:

• Bounded drivers: Hardware descriptions can define what an agent may read or change and the limits that the driver should enforce.
• Structured device information: Agents receive explicit details about capabilities and constraints instead of being asked to infer them from scratch.
• Deterministic execution for validated work: Reusable procedures can be compiled into fixed command sequences, reducing dependence on continuous model judgment.
• Limited early access: Partner testing gives Anthropic a setting to evaluate failures and establish operational practices before wider availability. 1

MHS can standardize communication and constrain commands, but it cannot by itself establish that an experimental plan is scientifically sound or that a lab is physically safe. Incorrect device descriptions, poor configuration, and unsafe procedures remain real risks. Human oversight, laboratory governance, and conventional physical safeguards therefore remain essential.

The bigger shift

CarPlay and MHS show two ends of the same strategy: bringing Claude into real-world workflows while limiting its authority to match the environment. In a moving car, that means voice conversation without control of the vehicle or phone. In a lab, it means controlled access to specialized hardware through standardized interfaces, declared limits, and carefully staged deployment. 171