Fifteen Equal Spokes, And Only One Of Them Clicks Things

This one is drawn as a sunburst: Claude in a white circle at the centre, fifteen numbered spokes radiating out, each ending in a labelled capability. Writing. Research. Data analysis. Artifacts. Claude Code. And so on round to Planning & Project Management.

As a map of what the product does, it is better than the task lists we usually get — capabilities age more slowly than use cases. The problem with a sunburst is that it makes everything equidistant. Every spoke is the same length, the same width, the same colour.

Fourteen of these fifteen hand you something to read. Number six moves a mouse.

Diagram: sort the fifteen by what comes back

The useful axis is not what each capability is called. It is what lands in front of you when it finishes, and whether anything changed while it ran.

Number six is not a thing you do. It is a thing you build.

The card’s line is “Let Claude use computers to click, type, navigate, and complete tasks for you.” Read next to “Brainstorming & Ideation”, that sounds like a setting. The documentation describes something with a different shape:

“Computer use requires a sandboxed computing environment where Claude can safely interact with applications and the web.”

“When you use computer use, Claude doesn’t directly connect to this environment. Instead, your application: receives Claude’s tool use requests, translates them into actions in your computing environment, captures the results… and returns these results to Claude.”

“For security and isolation, the reference implementation runs all of this inside a Docker container with appropriate port mappings for viewing and interacting with the environment.”

So the clicking is not done by Claude. It is done by a program you write, which receives a request that says “left click at these coordinates” and decides whether to carry it out. The model supplies intent and screenshots; the hands are yours to build and yours to restrain.

That is the difference between number six and its fourteen neighbours. Thirteen of them need a prompt. One needs a prompt and a project.

And it arrives with four precautions

The documentation opens the computer use page with a warning that none of the other fourteen capabilities carries:

“Computer use has unique risks distinct from standard API features. These risks are heightened when interacting with the internet.”

Then four numbered precautions, quoted here because they are the part a one-line card cannot hold:

1. “Using a dedicated virtual machine or container with minimal privileges to prevent direct system attacks or accidents.”
2. “Avoiding giving the model access to sensitive data, such as account login information, to prevent information theft.”
3. “Limiting internet access to an allowlist of domains to reduce exposure to malicious content.”
4. “Asking a human to confirm decisions that might result in meaningful real-world consequences and any tasks requiring affirmative consent, such as accepting cookies, completing financial transactions, or agreeing to terms of service.”

And the reason for all four, stated plainly:

“In some circumstances, Claude will follow commands found in content even when they conflict with your instructions. For example, instructions on webpages or contained in images might override your instructions or cause Claude to make mistakes.”

To be fair, there is a defence in the box as well. Anthropic says the model is trained to resist these injections, and that “classifiers will automatically scan what the tools return, such as screenshots, to flag potential prompt injections” and steer the model to check with you first. The same passage adds that “the precautions above remain important even with these classifiers in place”. Both halves are worth knowing.

The one label on the card, and where it is out of date

Credit first: of the fifteen, computer use is the only one the card flags, and the fine print names it again. That is the right item to single out. The label itself has drifted, though.

One more detail from the same page that explains why it matters more than a label: the current toolset gives Claude seventeen member tools — screenshot, left_click, type, zoom and the rest — from a single entry in the tools list. Fifteen capabilities on the card; seventeen separate ways to touch a machine inside one of them.

What else on the card needs something you do not have yet

Computer use is the extreme case, but it is not the only spoke with a prerequisite. Sorting the fifteen that way turns a poster into a plan.

What the card gets right

Three things. First, the framing: a capability map beats a task list, because tasks go stale and capabilities do not. “Analyse data, find patterns” will still read correctly next year; “build this specific app” will not.

Second, the fine print names exactly the right three: “Features such as Computer Use (Beta), connectors and Claude Code change often and may not be on every plan.” Those are precisely the three spokes on the card with a dependency outside the chat window. Whoever wrote that footer knew which rows were load-bearing.

Third, it does not overclaim about what the model is. There is no “replaces your team” line anywhere on it. Fifteen capabilities, one sentence each, no promises.

Four honest caveats

1. We did not build a computer use environment. Everything above is the card checked against the computer use tool documentation, in October 2026.
2. “Not a thing you do” is about the API route we quoted. Desktop and browser surfaces that drive a machine for you exist and are easier to switch on; our point is that the card’s one-line description covers a family of things whose easiest member still asks for more than a prompt.
3. Status labels move in both directions. A beta that is now GA can gain a new beta header next month. Check the page rather than the card, which is what the card’s own footer says too.
4. This is not an argument against computer use. It is an argument against drawing it the same size as “Multi-Language Support”. The capability is real and the documentation is unusually candid about its risks.

The one-line version

Fourteen of these fifteen spokes end in something you read; the fifteenth ends in a click, needs an application you write and a sandbox to run it in, and carries four written precautions — and it is drawn exactly the same size as “Brainstorming & Ideation”.

Start with the eight that need nothing. Treat number six as a project, not a feature.