claude code desktop can now browse the web — and /doctor can actually fix things

Claude Code's desktop app shipped a built-in browser in versions v2.1.202–206, released during week 28 (July 6–10, 2026). the feature does what it sounds like: Claude can now open a web page from inside your terminal session, read the content, click through links and buttons, and interact with external sites the same way it already interacts with your local dev server preview.

the practical use cases this unlocks are the ones that previously required you to manually copy-paste context into your session — pulling the specific section of docs that's relevant to what you're building, checking a design on Figma or a staging URL, verifying what an error page actually says vs. what the logs say it says. instead of you acting as the relay, Claude pulls the page itself.

how the sandboxing works

the browser is sandboxed and configurable. you choose whether browsing sessions persist between uses — meaning Claude can either start fresh each time or maintain state like cookies and session data across a session. Anthropic runs safety classifiers over Claude's actions on external sites before they execute. that's a different trust boundary than the one on your local files — external sites aren't under your control, and the classifier layer is how they handle that.

the documentation points to a sandboxed design rather than a fully open browser, which is the right call for an agentic tool: the failure mode of an unsandboxed agent clicking through external sites is meaningfully worse than one that's constrained. the configuration options give you the knobs to tune persistence vs. isolation based on your workflow.

/doctor now fixes things

the same release upgraded the /doctor command (now also aliased as /checkup) from a read-only diagnostic report into an interactive tool that can fix what it finds.

specifically, it now: checks installation health; identifies unused skills, MCP servers, and plugins and flags their context cost; deduplicates local CLAUDE.md files against checked-in versions; proposes trimming content Claude could derive from the codebase itself; and flags slow hooks. it shows you the findings first and asks for confirmation before changing anything — so you can audit before it acts.

the CLAUDE.md dedup and trimming proposals are the most useful piece for teams that have been using Claude Code for a while. context files accumulate cruft over months of sessions — rules that got superseded, documentation that now lives in the codebase, entries that conflict with each other. /doctor finding and proposing those trims is a genuine workflow improvement.

other w28 changes worth noting

a few smaller changes from the same release that are worth knowing about:

auto mode now blocks tampering with session transcript files, and asks before running rm -rfon a variable it can't resolve from context — a guard against the failure mode where an agentic loop deletes something it shouldn't because a path variable resolved unexpectedly.

background task notifications now explicitly state that no human input has occurred, which prevents fabricated in-transcript approvals from being acted on. this is a security-relevant change for teams running Claude Code in multi-agent or CI configurations.

the binary updater now streams downloads to disk instead of buffering them in memory — cutting peak memory usage during updates by roughly 400 MB. if you run Claude Code on machines with constrained RAM, that's a notable improvement.

agent view now shows a colored state label and a classifier-written headline instead of raw tool call text. if you've been monitoring background agent sessions and the output felt opaque, this makes the current state of a running agent more readable at a glance.