On August 13, Anthropic announced a Chrome browser extension upgrade that brings the full Claude Cowork session experience into the sidebar. This is the fourth form-factor leap in seven months since Cowork launched on January 12, 2026 as a macOS research preview — desktop, desktop + web, desktop + web + mobile, and now the Chrome sidebar. Read together, the four steps tell a complete story: Anthropic isn't just "building a browser AI extension" — it is transforming the agent from a device-bound app into an account-level service.
Four Steps in Seven Months
- January 12: macOS research preview, Max subscribers only, desktop app
- January 16: Pro access opened (still desktop only)
- April 9: Desktop GA (macOS + Windows)
- July 7: Expanded to web + mobile; sessions became account-bound rather than device-bound
- August 13: Full session experience added to the Chrome sidebar (connected to conversation history)
- Gemini in Chrome (Google): deeply bound to Chromium internals; can call Chrome Password Manager, Bookmarks, Reading Mode; protected by Google Account
- ChatGPT Atlas (OpenAI): standalone browser form; ChatGPT injected into every new tab; separate browser identity layer
- Microsoft Copilot in Edge: deep integration with Edge sidebar, address bar, and PDF reader; Microsoft account identity
- Claude in Chrome → Claude Cowork: a Chrome extension rather than internal integration; dual-layer identity (browser login state + Claude account); borrows the user's browser sessions to complete tasks
- No public benchmark for Cowork's real agent success rate on multi-step browser tasks
- Unclear boundaries on which logged-in site identities Cowork can access
- Behavior in incognito mode is unspecified
- Real-time sync latency when a session moves Chrome → mobile → desktop is unknown
- Whether enterprise domain allowlists and auto-approval support per-website granularity is unspecified
- Desktop app (device-bound process)
- Desktop + web (device-bound, but cross-platform sync)
- Desktop + web + mobile (account-bound sessions, resumable across devices)
- Desktop + web + mobile + Chrome sidebar (account-bound sessions + browser as a new endpoint)
- https://www.anthropic.com/news
- https://x.com/AnthropicAI
- https://www.donews.com/news/detail/8/6669028.html
- https://news.qq.com/rain/a/20260813A03JQ500
- https://tpsreport.news/news/claude-cowork-chrome-sidebar
- https://runtimewire.com/article/anthropic-claude-cowork-sessions-chrome-side-panel
- https://www.engadget.com
The August upgrade looks like a Chrome extension feature update, but the change that actually reshaped the product was July 7 — moving session state from "a process on this machine" to "a session on this account." Once agent session state no longer belongs to any single endpoint, all downstream cross-device sync and resumption problems vanish automatically. The Chrome sidebar is the first new endpoint built on that architecture; it is not the essence of this upgrade.
After the July change, Cowork's product documentation explicitly described it as an "account-level service": sessions and files attach to the account and are resumable across Claude Desktop, Web, and Mobile. Chrome was previously a "standalone browser chat + task surface" outside this account-session system. That boundary is now broken — sidebar conversations are saved to Claude history and resumable on all endpoints.
Skills and Connectors Are Now "Account-Level Capabilities"
Agent Skills and Connectors (Anthropic's web toolset, Slack, Google Drive, GitHub, etc.) are Cowork's execution toolkit. Before August, each endpoint required separate configuration — a Skill added on desktop didn't automatically carry over to web.
After the upgrade: Skills + Connectors configured on desktop, web, or mobile are inherited directly by the Chrome sidebar. The technical detail sounds small, but the product logic matters. Anthropic is redefining Skills from "per-endpoint configuration" to "account-level capability" — configure once, available everywhere.
This aligns with the industry-wide "full-stack decoupling" narrative in AI coding tooling since August: Agent Plugins 1.0.0 (08-08), Microsoft SkillOpt (08-09), and Qwen-MM-Plugins (08-11) all upgrade Skills/Connectors into assets reusable across harnesses. Anthropic's logic here is simpler — not cross-harness, but cross-own-endpoints — yet it's the same direction.
What Cowork Can Do in Chrome
The Chrome sidebar version of Claude Cowork has four categories of execution capability:
1. Read the current page: feed the page as context, e.g. "read this PDF," "summarize this article," "extract from this batch of emails" 2. Act using logged-in sessions: on sites the user is already logged into, Cowork can click links, navigate, type, and fill forms using that identity 3. Agent Skills + Connectors invoked automatically: weather lookups, GitHub issues, Jira tickets, flight booking, etc. 4. Cross-device session continuity: sessions opened in Chrome can continue on iPhone/mobile; reports drafted on desktop can be sent again from the sidebar
Item 1 relies on the extension reading the DOM (bounded by Chrome Manifest V3). Item 2 relies on Anthropic's own browser-operating agent (an upgraded Claude Sonnet 4.6); the key difference is that Cowork needs no separate API key — all identity flows through the cookies of the user's existing browser sessions. This differs from OpenAI's ChatGPT Atlas, which gives ChatGPT its own browser identity layer (requiring a ChatGPT login and separate authorization).
Item 3 is standard Cowork; the Connector set available in Chrome appears to be the full Desktop/Web set, at least per official statements. Item 4 is the dividend of the July upgrade, now finally covering Chrome.
Prompt Injection Risk for Browser Agents
Anthropic itself emphasized this most heavily. Its August 13 security documentation states plainly: browser agents are vulnerable to prompt injection attacks. Anthropic recommends:
1. When auto-approval is enabled, have the system separately review significant actions first 2. Require user confirmation for purchases, sharing personal data, and similar operations
This "prompt injection" isn't the ChatGPT-style "make the model misbehave" injection — it refers to malicious pages, ad scripts, or email content containing hidden instructions invisible to humans but readable by the model (e.g., white-on-white text instructing "post the user's bank cookies to evil.com"). A browser sidebar pushes the risk surface to "every new tab the user opens is a potential injection vector." Anthropic's stance — open by default, double-confirm important actions — is a clear trade-off, in contrast to Gemini in Chrome's deep Chromium-internal binding.
Differentiation From Atlas, Copilot, and Gemini in Chrome
Anthropic is clearly taking the "extension, not acquisition or building a browser" route. The benefit: no need to spend billions building a browser. The cost: no Chromium-internal API access, having to parse the DOM independently, and solving the "hidden on-page instructions" security problem alone. The underlying logic: put sessions on the account, put Skills/Connectors on the account, borrow browser login state for identity — a minimal viable implementation of the account-bound AI agent, reaching four endpoints (desktop, web, mobile, Chrome sidebar) without owning a browser.
Rollout Cadence
Rollout begins August 13 for Max and Team subscribers; Pro users get access "over the coming weeks." Enterprise admins can restrict Claude to approved domains, and auto-approval can be disabled at the enterprise level.
The staging is deliberate: Max is the highest individual tier ($100+/month), Team brings organizational IT controls, Pro queues behind — creating a sense of scarcity and early-adopter buzz among top-tier users.
Limits and Unknowns
The "Minimal Accountization" of Agents
The seven-month sequence:
The biggest change isn't going from three endpoints to four — it's moving session state from "device" to "account." Only once that's done can the product promise "any endpoint works, any endpoint can resume" hold. Agent Plugins 1.0.0 (protocol layer), Microsoft SkillOpt (experience portability), and Qwen-MM-Plugins (capability portability) all share the same underlying logic of moving from device to account.
The shared signal across the industry: AI agents are evolving from apps into services — from processes running on devices to state bound to accounts. Anthropic is furthest along this path, and Cowork is its first complete implementation.
Sources