[FEATURE] Browser view / browser tools in Claude Code cloud sessions (Claude Code on the web)

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Browser view (the Claude in Chrome integration) only works in local sessions. Cloud sessions on Claude Code on the web run in headless Anthropic-managed VMs with no browser, so Claude can't visually verify frontend work: no screenshots, no clicking through a flow, no reading console errors from a rendered page.

This breaks the main promise of cloud sessions for frontend tasks. I can delegate "fix the login form validation" to a cloud session, but Claude can't open the app and confirm the fix works. I end up reviewing the PR blind or pulling the branch locally, which defeats the point of delegating.

The gap is sharpest away from my desk. Cloud sessions are the surface I reach for from my phone. My laptop may be nearby, but it's in a bag with the lid closed and no guaranteed connectivity. That's precisely the fire-and-forget scenario cloud sessions are designed for, and it's exactly where browser verification is unavailable.

The DIY route inside the sandbox is also closed, per existing bug reports:

• #11791 — Playwright/Puppeteer/Selenium can't run because the sandbox proxy doesn't support HTTPS CONNECT tunneling
• #15583 — Chromium binaries can't even be downloaded; Playwright CDN domains aren't on the egress allowlist
• #73564 — headless Chromium gets ERR_CONNECTION_RESET on all sites even with network access set to Full

This request is the feature-level ask that would resolve what those three report as symptoms.

It's also a widening gap against other cloud coding agents. GitHub's Copilot coding agent has shipped a built-in browser (Playwright MCP) with screenshots in the pull request since July 2025, roughly a year now. Cursor's cloud agents run a full desktop with a real browser and attach screenshots and video to the PR (February 2026). OpenAI's Codex cloud can surface UI screenshots for frontend tasks (September 2025). Claude Code cloud sessions are the one surface among these with no browser verification at all, and the localhost-scoped version below is a smaller, safer step than what those competitors already ship.

Proposed Solution

First-class browser tools in cloud sessions, scoped to the sandbox first:

1. Pre-install headless Chromium, opt-in per environment config, the same way packages and networking are configured today. Sessions that don't enable it pay nothing in image size, memory, or cold-start time. This removes the download blocker (#15583) without adding weight to the default VM.
2. Allow the browser to reach sandbox-internal addresses (localhost) without going through the egress proxy. The primary use case is Claude starting the dev server it just modified and verifying it: navigate, click, type, read console, screenshot. Localhost traffic never leaves the VM, so this sidesteps the CONNECT tunneling limitation (#11791) and adds no new egress surface. It also adds no meaningful prompt injection surface: the browser renders the user's own repo code, which Claude can already read as files in the same session. Chromium's own external calls (update checks, safe browsing) are disabled by standard headless-shell builds and launch flags, as CI runners have done for years.
3. Expose this as browser tools with screenshots surfaced in the session transcript, visible in the claude.ai/code web view and mobile app, so I can see what Claude saw when reviewing from my phone.

External site browsing through the proxy is a separate, harder problem (CONNECT support has real security trade-offs for an inspecting proxy). It could be phase 2, opt-in per environment config the same way networking and packages are configured today. The localhost-only version already delivers most of the value.

Alternative Solutions

Remote Control is the current answer for browser-verified work away from the desk, and it's fragile in exactly the situations where cloud sessions shine:

• It requires my laptop to be open, awake, online, and running an active Claude Code process for the entire session. Having the laptop with me isn't enough; it has to be in active service the whole time, which defeats the point of steering from a phone.
• The session dies after roughly 10 minutes of network disruption on the host machine, so it can't be left running unattended.
• It's a single active session mirrored to one device, not the parallel fire-and-forget model cloud sessions offer.

Headless browser via bash in the sandbox is blocked at three layers (binary download, proxy tunneling, connection resets) per the issues above, and even if it worked it gives Claude no first-class screenshot/console tooling.

Pull the branch locally to verify works but removes the value of delegating in the first place.

Priority

High - Significant impact on productivity