Skip to content
Agentsmedium signalverified

Claude gets a separate toolset for driving a browser

August 19 brought two changes to how Claude operates someone else's screen. The computer use tool moved to general availability as computer_toolset_20260801, with no beta header, actions grouped into a single turn, and zoom enabled from the start. Alongside it came browser_toolset_20260801, a separate set of 31 tools that works inside a browser you run yourself rather than across the whole desktop. The important difference is that this set reads the page accessibility tree and returns element references, so a click targets a reference such as ref_2 instead of a point in pixels — and that survives a layout shift.

By Redakcija WebAiRadarPublished 3 min readwritten by a model

Source

Browser use tool

Claude dokumentacija · Original published August 19, 2026

On August 19, Anthropic split two jobs that had been going through the same tool. Driving the whole desktop stayed in computer_toolset_20260801, which left beta the same day, while working inside a page got its own browser_toolset_20260801. The difference is not cosmetic: the new set sees the structure of a page, not just a picture of it.

Thirty-one tools, sorted by the job

The set ships with 27 tools enabled from the start and four more you turn on yourself. The split follows what people actually do in a browser.

  • Navigation and capture: navigate, screenshot, and zoom.
  • Pointer: clicks with every button, double and triple click, hover, drag, separate press and release, movement, and scrolling.
  • Keyboard and timing: type, key, hold_key, and wait.
  • Reading the page: read_page returns the accessibility tree with references, find locates an element from a plain-language description, and get_page_text extracts visible text, favoring the main content.
  • Forms and tabs: form_input sets a field value directly, while new_tab, list_tabs, switch_tab, and close_tab manage tabs.

An element reference instead of a point in pixels

Actions take a target in one of two forms. The first is a viewport coordinate, x and y, the same shape the computer use tool understands. The second is an element reference such as ref_2, returned by read_page and find.

The second form is what makes this set usable on real sites. A reference survives a layout shift and requires no pixel guessing, so a button that moved after an image loaded is still hit. References hold until the tab navigates to a new address or the page structure changes substantially, and they belong to a single tab.

Browser state comes back with the result

Results can carry a browser_state block, in which the executor reports what actually happened in the browser: which tabs are open, with title and address, which one is active, whether any opened or closed, and whether a download started, finished, or failed.

  • A single block holds at most 100 tabs, and exactly one carries the active marker.
  • A tab reference is a stable string of at most 4,096 characters, with no control characters.
  • Several actions in one turn run in order, but after the first failure the rest do not run and a message says so.

What stays off until you turn it on

Four tools ship disabled, and for good reason. Running JavaScript in the page and uploading files give the most power, while reading the console and network traffic serve diagnostics. You enable them individually through the configs array, and only once your executor actually supports them and the task calls for them.

Where it runs and where it does not

  • Supported models are Claude Fable 5, Claude Mythos 5, Claude Opus 5, Claude Sonnet 5, and Claude Opus 4.8.
  • The set is available only on the Claude API. It is not on Amazon Bedrock, Google Cloud, or Microsoft Foundry, and not within Managed Agents.
  • There is no separate billing. Usage counts as ordinary token consumption, with the caveat that page reads and screenshots consume input context.
  • The browser runs in your environment, not on Anthropic's servers.

Security is a condition here, not a recommendation

The documentation is explicit on this point, because executing someone else's content and taking real actions meet here. The browser should run in an isolated environment without credentials, the allowed-domain list belongs on the network, and navigate should accept only http and https addresses. Actions with consequences, such as a purchase or an account change, require a human to confirm.

Pages can contain prompt injection text.
Anthropic, browser tool documentation

Sources

BrandsClaude

Related

Agentsmedium signal

A proxy that stripped one header was doubling Claude Code's API bill

Version 2.1.239 fixes streaming on Bedrock behind proxies that remove the response Content-Type header. Claude Code silently fell back to re-running every turn without streaming, and each turn was billed twice. The same release makes cost estimates show the 1.1× premium that data-residency workspaces pay.

Anthropicverified

CHATGPT FOR MACiMessagereads and sends, only after you approve
Agentsmedium signal

ChatGPT now reads and sends your Apple Messages, on the Mac only

A plug-in in the ChatGPT desktop app reads and searches your iMessage, SMS and RCS threads and sends messages through Messages on your behalf. By default nothing goes out until you approve both the text and the recipients. There is a switch that removes that step, and OpenAI's own documentation argues against using it.

OpenAIverified

Screenshot of the GitHub Copilot panel inside Slack, under the heading "The new GitHub Copilot experience in Slack".
Agentsmedium signal

GitHub Copilot moved into Slack and Microsoft Teams on the same day

Both shipped on August 21 in public preview. Mention @GitHub in a channel and the agent triages issues, investigates failures, writes changes in a cloud sandbox and opens a pull request, with the conversation attached. The interesting part is not the capability list, which is familiar, but the room it moved into: the place where work gets discussed rather than written.

GitHubverified