Anthropic · Claude Code
The terminal AI has started
opening its own browser.
Claude Code now ships with a built-in browser. Browser control that used to require wiring in a Playwright / Puppeteer MCP plugin now works inside Claude Code by itself. Post-login admin screens, form fills on internal tools, deep documentation lookups — the "let me just open the page and check" step can now be closed by the AI. THE DECODER broke the news first.
The Change
Browser is now a feature,
not a plugin
The line moved from "attach a browser via MCP" to "there is a browser inside Claude Code."
Until now, letting Claude Code touch a real page meant installing an MCP plugin that drove Playwright or Puppeteer. It worked on your machine, but carrying auth sessions across processes, reproducing them in CI, and reconciling Windows / macOS differences ate into initial-setup time.
Bringing the browser inside removes the class of "should work but the plugin isn't installed" failures. From the CLI you invoke browse like any command; Claude opens the page, clicks the button, fills the form, and reads the result — all inside one session. THE DECODER's first line, "AI can open sites, click, and fill forms itself, no MCP plugin needed," captures the change concisely.
By The Numbers
Six practical facts
Why It Matters
What "closing the loop" really costs and buys
This is less a feature bump and more a boundary shift — coding AIs move from strings to screens.
Coding AIs used to stop at "write code, run tests, open a PR." Actually seeing whether the running thing looks right was the human's job. MCP plugins narrowed the gap, but the overhead pushed people to skip them for anything not clearly worth it. With the browser inside Claude Code, the implement → verify → fix loop can now close inside the AI.
Individually, the win is small — a few minutes of manual clicking gone. Where those few minutes recur (PoCs, E2E checks, admin-panel sweeps), the daily rhythm changes. Codex and Cursor remain on the MCP model, so Claude's "bundled browser" edge could hold for months.
Who It Hits
Who feels this, and how
Front-end / app dev
Have Claude open the screen you just built and walk it through the form. First-pass manual QA now finishes inside Claude Code.
Internal-tool ops
Delegate small admin-panel jobs (delete a user, flip a flag, confirm a change). Bastion-host clicking shrinks.
Deep doc research
Let Claude walk deep documentation trees to build answers with citations. Deeper than "paste a URL, summarize."
Coding AI's boundary just moved
from strings to screens.
What To Do Next
Three experiments for this week
Replace one manual QA scenario with the built-in browser
Take the "sign up → first login → home renders" path your QA always runs first and drive it from Claude Code. Turning those few minutes into "it just runs when I implement" protects developer focus.
Trim the MCP plugin list
Browser MCPs can stay for now, but drop the ones "the built-in already covers". CI setup lightens up. The audit is a one-week job.
Re-check your permission perimeter
More reachable pages means more credential exposure and more chances of unintended writes on production admin screens. Write down which URLs, domains, and roles Claude Code may touch this week.
Counterview
Counterview and limits
An optimistic reading is only half the picture. Letting AI touch production admin panels means handing write permissions to Claude Code — and now that the browser is bundled, indirect prompt injection is easier to overlook. Keep human approval in front of write-side actions this week too.
Second, "bundled" means "maintained by Anthropic on their schedule." Similar in-CLI features have been externalized months later. Whether today's convenience becomes a long-term design assumption is not settled yet. For CI, keep a Playwright fallback path warm.