When Will Anthropic Model Hardware Standard Open? 2026 Preparation Checklist

A team waiting for MHS may be delaying basic device-security work that it can already complete.

The fastest answer: as of September 10, 2026, Anthropic Model Hardware Standard has no confirmed full open-source date; prepare your device APIs, permission boundaries, emergency stop controls, and replayable tests now, but do not redesign around an unannounced launch.

This article is for:

• Developers who want to know when MHS may become usable in ordinary Agent projects.
• Research and manufacturing teams operating programmable instruments, robots, or remote equipment.
• Platform owners planning device-Agent safety validation, remote control, and compute environments.

Status check: Anthropic has confirmed a research preview for selected research labs and advanced manufacturing organizations, with waitlist applications. Anthropic has said that open sourcing is planned, but it has not published a specific date or a complete public compatibility list. See the official MHS research preview announcement for the current primary source.

The opening-time answer and the evidence boundary

The most important distinction is between a research preview and broad developer access. A preview can allow Anthropic to test device abstractions, safety controls, and operational assumptions with selected organizations. It does not mean that any developer can download a complete specification, connect an arbitrary instrument, or claim official compatibility.

As of September 10, 2026, the confirmed position is narrow:

• MHS is being offered to an initial group of research laboratories and advanced manufacturing organizations.
• Applications are being accepted through a waitlist process.
• Anthropic has stated that an open-source release is planned.
• No specific full opening date has been announced.
• Any claim about a particular quarter, expanded device list, or universal availability remains a forecast rather than a confirmed roadmap item.

That makes the correct planning decision conditional. If your project depends on a public MHS specification, wait for the official release before committing to protocol-specific implementation. If your project depends on safe Agent control of physical equipment, do not wait. The surrounding engineering work remains useful whether MHS arrives soon, changes shape, or stays limited to selected preview partners.

The phrase “MHS opening time” therefore needs two meanings. It can refer to access to the preview, which is already limited and application-based. It can also refer to broad access to an open specification and usable implementation guidance, which has not received a confirmed date.

Software Agents and desktop automation

MHS is not a general label for every computer task performed by Claude or another AI Agent. A desktop workflow that opens applications, edits files, calls APIs, or runs scripts is not automatically a hardware-standard use case. The same applies to a normal MCP server that exposes a software tool.

For these teams, the immediate risks are different:

• A tool may have broader write access than the task requires.
• Credentials may be available to an Agent process for longer than necessary.
• A desktop session may mix test data with personal or production accounts.
• Operators may be unable to reconstruct the exact sequence after a failed action.
• A long-running task may continue after its business approval has expired.

The Anthropic MCP documentation describes MCP as a way for AI applications to connect with external tools and data sources. That is an integration layer, not proof that a connected device is covered by MHS. Your architecture should keep the boundaries explicit:

• The model decides or proposes.
• The Agent harness manages task state and tool selection.
• MCP exposes a software capability.
• A device driver translates commands into equipment-specific operations.
• A safety layer checks limits, approval, interlocks, and emergency actions.
• The physical device remains responsible for its own low-level protective controls.

This separation is more valuable than waiting for a protocol name. A desktop automation team can start by isolating the operating environment, using short-lived credentials, recording every tool call, and requiring approval for destructive operations. A future MHS adapter can then sit behind the same policy layer rather than receiving unrestricted access to the host.

For a computer-control project, use a remote Mac environment as a reference point when evaluating isolated sessions and repeatable Agent workflows. The relevant question is not whether the environment is officially MHS-compatible. It is whether you can reproduce a task, inspect its permissions, and stop it without affecting unrelated systems.

Laboratory instruments and experimental automation

Laboratory automation is closer to the intended MHS problem because the Agent may need to interact with programmable physical equipment. However, “closer” does not mean “already supported.” You should not describe an existing spectrometer, controller, microscope, pump, or test fixture as MHS-compatible unless Anthropic publishes that status or your organization receives explicit preview confirmation.

The preparation work is concrete. Build an equipment profile for each device that records:

• The programming interface and transport method.
• Readable state, telemetry, and calibration information.
• Valid command ranges and units.
• Error codes, timeout behavior, and recovery actions.
• The difference between a reversible command and a hazardous command.
• The location and behavior of the physical emergency stop.
• The conditions that require a human operator to take control.
• The evidence required before an experiment can resume.

The Agent should not infer safety from a successful API response. A command can be accepted by a driver while the instrument is in the wrong physical state, a sample is missing, a lid is open, or a downstream system is unavailable. Your validation layer needs state checks that are independent of the language model.

Anthropic’s published work on Claude and robotics is useful for understanding why language-level planning is not equivalent to safe physical execution. The Claude robotics research shows the relevance of model-driven interaction with robotic tasks, but it does not turn your own laboratory equipment into a confirmed MHS target.

A research laboratory should also prepare a human-supervision protocol. Define who can approve a run, who can pause it, who can restart equipment, and which events require a manual inspection. Store the device state before and after every action so that a reviewer can distinguish an Agent planning error from a driver fault or a hardware fault.

Robotics and coordinated devices

Robotics teams may gain the most from a common device interface if it eventually supports consistent discovery, state reporting, command validation, and cross-vendor orchestration. A task planner could theoretically coordinate a robot arm, vision system, gripper, conveyor, and inspection station through a clearer abstraction than a collection of unrelated vendor APIs.

That potential should remain a roadmap implication, not a current capability claim. The difficult work is not only sending commands. It is understanding space, timing, force, visibility, and failure recovery.

Prepare the following boundaries before considering MHS integration:

• Maximum permitted movement and speed.
• Forbidden zones and collision envelopes.
• Grip, load, temperature, and force limits.
• Sensor disagreement handling.
• Loss-of-connection behavior.
• Local and remote emergency-stop paths.
• Manual takeover and hand-back procedures.
• Simulation tests for expected and unexpected states.

The NIST physical AI and robotics project provides a useful public reference for thinking about data generation and evaluation in physical systems. It does not certify MHS devices, but it reinforces the need to test physical behavior rather than judge an Agent from text output alone.

Spatial reasoning also needs expert supervision. A robot can follow a syntactically valid plan while the workspace has changed. A camera can return stale data. A gripper can report a command success without holding the object securely. Your test harness should therefore include sensor loss, delayed state updates, unexpected obstacles, and failed actuators. Those cases are valuable before any standard becomes available.

Manufacturing and remote equipment operations

In manufacturing, MHS could eventually make device discovery and multi-machine coordination easier. A future interface might help an Agent identify available equipment, read operating states, request a recovery action, or coordinate a sequence across machines. These are possible effects on the roadmap, not promises about the current preview.

The immediate operational plan should be more conservative:

• Keep monitoring read-only by default.
• Put every write operation behind an approval gate.
• Separate the remote control network from general office access.
• Require local confirmation for hazardous movement or energy changes.
• Record operator identity, requested action, device response, and final state.
• Test recovery with simulated faults before touching production equipment.

Remote control creates a second risk beyond the physical machine: the network path can fail while the command state remains ambiguous. A safe design must define whether a timeout means “retry,” “stop,” or “inspect manually.” Do not allow a general-purpose Agent to decide that policy during an incident.

Use the NIST AI Risk Management Framework as a governance reference. Its four core functions, Govern, Map, Measure, and Manage, give platform teams a way to place physical-device risks into an existing review process. Those four functions are not an MHS requirement, but they provide a useful structure for documenting ownership and evidence.

A manufacturing team can also create an interface adapter now. Keep the adapter responsible for device discovery, state normalization, command validation, and error mapping. Keep policy decisions outside the adapter. This design lets you replace a vendor driver or add a future MHS connector without rewriting the entire Agent workflow.

Agent platforms, MCP, and device drivers

MHS and MCP should not be treated as competing versions of the same thing. MCP is a protocol for connecting AI applications with tools and data. The MCP announcement explains that broader integration goal. MHS, based on the confirmed preview description, is aimed at programmable physical hardware and the interfaces needed to operate it safely.

The relationship is more practical than semantic:

• MCP may expose a software-facing tool to the Agent.
• MHS may define or standardize how a physical device presents capabilities and state.
• A bridge may translate MHS device operations into MCP tools.
• Your authorization system must still decide which Agent can call which operation.
• Physical interlocks and emergency controls must remain outside the model’s authority.

The MCP authorization specification is dated June 18, 2025 and documents an OAuth-based authorization model for protected MCP resources. That date is a useful reminder that protocol specifications evolve independently. Pin versions, record the authorization behavior used by each connector, and do not assume that a future MHS release will preserve your current tool schema.

For a platform team, the durable components are credentials, policy, logs, replay, test nodes, and version management. The model should be replaceable. The Agent harness should be testable without a live machine. Device drivers should expose typed states and bounded actions. A failed request should produce an inspectable event, not only a natural-language explanation.

Configuration comparison

The comparison points to a clear investment rule: build controls and adapters that remain useful under several protocol outcomes, but avoid implementing unverified MHS-specific assumptions.

The preparation checklist

Use this checklist before applying for a preview or connecting any physical device to an Agent:

• [ ] Record every device API, transport method, supported operation, and readable state.
• [ ] Separate read-only telemetry from commands that change physical conditions.
• [ ] Define approval requirements for each write operation.
• [ ] Document error codes, timeout behavior, retry limits, and manual recovery.
• [ ] Identify the physical emergency stop and test who can activate it.
• [ ] Create a sandbox or simulator that does not control production equipment.
• [ ] Capture model version, Agent version, tool version, device state, and operator identity in each run.
• [ ] Build replay cases for normal completion, stale sensors, lost connections, invalid commands, and operator takeover.
• [ ] Keep credentials in a managed store rather than in prompts, scripts, or shared environment files.
• [ ] Pin MCP and driver versions so protocol changes can be detected during testing.
• [ ] Define the boundary between model reasoning, Agent orchestration, MCP tools, and device safety controls.
• [ ] Assign a human owner for each device class and each emergency procedure.
• [ ] Create an evidence folder for preview applications, official guidance, test results, and approval decisions.

This is also the answer to what a device Agent should prepare before MHS access. You need a controlled interface and evidence trail, not only a working demo.

The update board for the waiting period

Track MHS work in three separate columns so that speculation does not become an engineering commitment.

Actionable now

• Normalize device capabilities and states.
• Build permission policies and approval flows.
• Add simulation and replay.
• Test remote isolation and emergency procedures.
• Review desktop Agent and MCP security independently of MHS.

Waiting for preview eligibility

• Application acceptance.
• Access to preview documentation.
• Confirmation that your equipment category is in scope.
• Feedback on the proposed device interface and safety process.

Must wait for a public specification

• Claims of broad compatibility.
• Production dependence on an unannounced schema.
• A public support matrix.
• Forecasts about a universal opening date.
• Large-scale rewrites based on partner statements or media predictions.

For ongoing review, check Anthropic’s Newsroom, the official MHS page, any public specification repository, supported-device guidance, and safety documentation each week. Compare the wording of application status, open-source plans, supported protocols, and safety requirements line by line. Do not infer an official release date from a partner announcement.

When an official announcement changes the status, update this page before creating another short news item. That keeps the timeline, preparation advice, and evidence boundary in one place.

Choosing a temporary test environment

A remote Mac can help with desktop Agent isolation, tool-permission testing, task replay, and audit workflows. It cannot prove that MHS officially supports a physical instrument or robot. This distinction matters because a reproducible computer environment tests the control plane, while physical safety also depends on device interlocks, sensors, workspace conditions, and trained operators.

For teams without a stable test host, review SSHMac’s environment and support guidance before selecting a temporary setup. Evaluate the environment against your own acceptance criteria:

• Can you create a clean session for each test?
• Can you restrict credentials and network access?
• Can you retain logs after the session ends?
• Can another engineer replay the same workflow?
• Can a human stop the Agent without relying on the model?
• Can you keep simulated device tools separate from production equipment?

If the answer is no, fix that control gap before spending engineering time on a speculative MHS adapter.

The practical route from current tools to future hardware standards

Your current desktop Agent stack may be fast to deploy, but it often has weaker physical boundaries, inconsistent audit records, shared credentials, and limited recovery testing. Direct vendor integrations can also create one-off drivers that are difficult to replace. A cloud-only workflow may simplify access while leaving network isolation, session persistence, and emergency ownership unclear.

That is why SSHMac can be useful as a temporary Mac-based testing option when you need to validate Agent isolation, permissions, and repeatable automation before a physical-device pilot. It is not a substitute for MHS, a robotics safety system, or official Anthropic compatibility. It is a way to test the surrounding control plane without turning an uncertain roadmap into a production dependency.

The right decision for now is simple: wait for the confirmed MHS opening date before building around the standard itself, but use the waiting period to make your Agent environment auditable, bounded, replayable, and safe to hand over to a human.