Last week was governance. This week is surface area. The Claude in Chrome side panel stopped being its own separate thing and became a Cowork session, which means a task you start in a browser tab can be finished on your desktop. Claude Tag learned to read a whole Slack channel instead of one message at a time. And Anthropic published a long explainer on the text watermark that future Claude models will carry.
The Chrome Side Panel Is Now a Cowork Session
On August 12, Anthropic announced that the Claude in Chrome side panel now runs the same Claude Cowork session you use on desktop, web, and mobile. Practically, three things change: conversations in the side panel are saved to your history instead of disappearing, your skills and connectors work inside the browser, and a task you start in a tab can be picked up later somewhere else. Because sessions live with your account rather than a single device, the browser stops being a dead end.
The reason this matters is the gap it closes. Plenty of tools connect to Claude directly through connectors, but internal dashboards, legacy systems, and vendor portals do not - and probably never will. Claude in Chrome reaches those through the browser using your existing logins. Anthropic's own example: ask Claude in Chrome to pull invoice amounts and dates out of several vendor portals and build a spreadsheet, then open that same session in the desktop app to add files from your computer and compare against last month's budget.
It is available on Max and Team plans today and rolling out to Pro over the coming weeks. On Enterprise plans, Claude in Chrome is off by default; admins turn it on and can limit it to approved domains. Two limits worth knowing up front: you still need the desktop app to work with files on your computer or with other applications, and the extension does not run on other Chromium browsers or on mobile yet.
A Second Check on Claude's Own Actions
Anthropic is direct about the risk: an agent acting in a browser is exposed to prompt injection, where instructions hidden in a page, an email, or a document redirect Claude toward something you never asked for. Since the pilot, they have added a check on Claude's own actions. If you turn on automatic approval so Claude works through a task without stopping at every step, a separate check reviews anything consequential - submitting a form, sending a message, downloading a file - against what you originally asked for, and blocks whatever does not match.
Claude still stops and asks before certain irreversible or costly actions, like making a purchase or sharing personal data. Anthropic is careful to note that these measures meaningfully reduce the risk but cannot eliminate it, and that prompt injection is a moving target. Their recommendation is to start on sites you trust.
Claude Tag Reads the Whole Channel Now
On August 13, Claude Tag - the Slack integration for Team and Enterprise plans - got a rework of how it decides when to speak. Previously a lightweight classifier looked at each new message on its own and made a yes-or-no call. That classifier is gone. Claude now uses context from across the channel, plus its memory and any standing instructions you have given it, to pick one of four moves: reply inline, start deeper work in a thread, route the message into a workstream it already has open, or say nothing.
Anthropic reports Claude is roughly 30% better at determining when, and when not to, proactively respond. The extra context it holds does not count toward usage or spend limits on any plan. It also knows when to stop paying attention: in a channel where it keeps concluding it has nothing to add, it goes to sleep, and an @-mention wakes it instantly. You can steer it in plain language - never respond here unless someone tags you - and any member can switch automatic responses off entirely.
Future Claude Models Will Watermark Their Text
On August 14, Anthropic published an explainer on text watermarking. Future Claude models will generate text carrying a watermark - a statistical pattern that lets someone with the key estimate the likelihood Claude was involved in writing it. This is EU AI Act compliance: Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside around 190 total signatories, and other major model developers are implementing their own watermarks. Anthropic is applying it globally at launch because they do not yet have a durable way to scope it by region.
The method is a version of Google DeepMind's SynthID-Text approach, published in Nature in 2024. It works by changing the source of the randomness Claude uses when picking between words that are equally good. Nothing is added to the text, there are no hidden characters, it requires no extra tokens, and it does not change the price or the quality of outputs. The watermark carries no identifying information and cannot be traced to a person, organization, or chat.
It has real limits, and Anthropic lists them plainly. Detection works poorly on short passages. It is sparser on factual writing, where there are fewer arbitrary word choices to hide a pattern in, and on code, which usually has to be exact - though comments inside code can carry it. If you hand Claude your own writing and ask only for grammar fixes, there may be too few changed words to register at all. And a complete rewrite removes it. For files, Claude attaches a C2PA content credential in the metadata of supported types like .png, .jpg, and .svg instead - a signed note saying the file was made or processed with Claude, which changes nothing inside the file. A watermark detection API is coming; the details are still being worked out.
Why This Matters for Cowork Users
The through-line this week is that a Cowork session is becoming something that lives in your account rather than in one app. It started as a desktop feature, went to web and mobile in July, and now includes the Chrome side panel. The practical upshot is that you should stop thinking about where to start a task and start thinking about which surface can reach the thing you need - browser for portals and internal dashboards you can only log into, desktop for files on your machine - because you can move between them without losing the thread.
Tip of the Week: Pick the Surface, Then Hand Off
Now that the side panel and the desktop app share one session, the useful habit is to split a task by what each surface can actually reach, and hand off in the middle - rather than fighting the wrong surface for the whole job.
❌ Weak
Start in the desktop app and ask Claude to get numbers out of a vendor portal it has no connector for, then re-explain the whole task in the browser when that fails.✅ Strong
In the Chrome side panel: "Pull the invoice amounts and dates from these three vendor tabs and build a table."
Then open the same session on desktop: "Now compare that table against last month's budget file in this folder."The rule of thumb: the browser gets you into things you can only log into; the desktop app gets you into files on your machine. One session, two reaches.
Worth Reading This Week
- → The Claude in Chrome side panel is now Claude Cowork - Claude Blog - The August 12 announcement, including plan availability, the action-check safeguard, and what the extension still cannot do.
- → Claude Tag now reads even more of the room - Claude Blog - The August 13 update on channel-wide context, the four moves, and how to steer or silence it.
- → How Claude's text watermark works - Anthropic - The August 14 explainer: the SynthID-Text method, what it can and cannot prove, and C2PA credentials for files.
- → Claude release notes - Anthropic Support - The running log of what shipped, worth checking before you assume a feature is or is not available on your plan.
Cowork keeps getting more places to live. The thing worth tracking is not each new surface on its own, but the fact that they now share one session - which quietly changes how you should plan a multi-step task.
Until next time.
— The ClaudeCowork.com team