OnlyRead
Star102

How do you keep Claude in Chrome read-only on LinkedIn?

A browser window showing a networking site: a profile card and a message thread with a green check, while a pink cursor is stopped by a pane of glass sealed with a pink padlock in front of the send and connect buttons.

The short answer

Even on Manually approve, its strictest setting, Claude in Chrome only asks you to OK the websites it will use and a rough outline of the task, then works on them on its own. Once you’ve said yes to LinkedIn, only Claude’s own judgment decides whether it asks before it sends a message, posts, comments or sends a connection request. Some apps let you connect an AI assistant with read-only access, but LinkedIn doesn’t. So on LinkedIn, and on any site where a slip would hurt, add OnlyRead, which refuses any request that would change something, whatever Claude decided.

You want Claude in Chrome to read a long thread in your LinkedIn messages, or go through the profiles of people who applied to a role, and nothing else. It shouldn’t reply to anyone, like a post, leave a comment or send a connection request along the way. Claude’s own settings help, and you should use them. But every one of them runs inside Claude: Claude decides what it will do, when a step needs your OK and what looks risky. When Claude misreads the task, that same misreading shapes what it asks you about. So on the sites that matter, you should add a second check that doesn’t depend on the agent at all.

Why does LinkedIn need the browser?

Some apps let you connect an AI assistant with read-only access, so it can search and read without opening a tab and can’t send anything. LinkedIn has no such option. Claude has no LinkedIn connection, and the messaging tools LinkedIn offers outside apps are for sending, open only to approved partners.1 There’s no read-only connection you can turn on for Claude. Unofficial LinkedIn add-ons exist, but many simply sign in as you, with everything your account can do. That leaves Claude in Chrome clicking through the real site, next to the same Message, Post, Comment and Connect buttons you use.

Which Claude in Chrome setting should you use for reading?

Claude in Chrome has three permission modes, chosen from the drop-down on the chat input.2 For work that should only read, choose Manually approve, but know what it approves. Anthropic’s summary of the mode says Claude asks before each action, but in the classic panel Claude first shows you the websites it will use, you approve them, and after that it works on those sites without asking at each step.

The labels below come from Anthropic’s help pages as of November 17, 2026. They describe the classic Claude in Chrome side panel. If the panel’s three-dot menu offers Switch back to classic, you’re in the newer Cowork panel, which asks differently.3

ModeWhat Claude doesFor a read-only job
Manually approve (was “Ask before acting”)Shows you the websites it will use, then works on them once you approve.Choose this, and check the list before you approve it
Automatically approveReviews each action for safety, blocks unsafe ones and asks you when needed.Workable, but Claude decides which actions reach you
Skip all approvals (was “Act without asking”)Doesn’t pause, and nothing checks its actions automatically.Avoid on accounts that matter

In Automatically approve, Anthropic says, Claude does the actions it judges lower-risk “and blocks or stops for anything that looks unsafe”.4 That judgment is Claude’s own, so a LinkedIn reply that looks routine to Claude can go ahead without a prompt, even when you only asked it to read. Anthropic’s permissions guide counts “messages sent as you” among the work “with real consequences” and advises you to “stay close and review what Claude does or consider switching back to ‘Manually approve.’”2

What Manually approve actually approves

In Manually approve, Claude turns your request into a short outline of the websites it will use and how it will go about the task, and you choose Approve plan or Make changes. Once you approve, Anthropic says, Claude “will be able to act independently within the outlined parameters”.2 It asks again to use a website you didn’t approve, for a sensitive action such as downloading a file, or when it judges that a step goes beyond what you approved. That last check is up to Claude. Nothing else asks before Claude clicks Send on a LinkedIn message, posts, comments or sends a connection request.

If the outline mentions replying, commenting or accepting invitations, choose Make changes and say it should only read. Approving “read my LinkedIn messages” still puts LinkedIn in Claude’s hands, though, and a misread step on an approved site goes through.

Remove site approvals you no longer need

When Claude needs a website you didn’t approve, it asks with Allow this action, Always allow actions on this site or Decline. An “always allow” lasts beyond today’s task. To review them, click the Claude extension icon, then the three dots at the top right of the side panel, then Extension settings. The Permissions page lists Your approved sites, and you can revoke any of them there.2 On an “always allow” site, Claude can take most actions without asking, sends and replies included. It still asks before downloads, entering sensitive information, granting authorizations, changing site permissions and permanent deletions.

What does Claude already ask about, or refuse?

Anthropic’s permissions guide lists actions that need your explicit permission in every mode:2

  • Changing permission settings
  • Granting authorizations
  • Entering potentially sensitive information into websites
  • Permanent deletions, such as emptying the trash or deleting emails, files or messages

It also lists actions Claude won’t take whatever the permissions, among them purchases and financial transactions, creating accounts, handling credit card or ID data, downloading files from untrusted sources and modifying system files.

Neither list includes sending a message, posting, commenting, reacting, sending a connection request or archiving a conversation. Those are the everyday changes a reading task can slip into. Once you’ve approved LinkedIn, Manually approve doesn’t stop to ask about them unless Claude judges they go beyond what you approved, and in Automatically approve it’s Claude’s call. Anthropic’s permissions guide warns that on a site you always allow, “Claude may take unintended actions across the website”.2

Why add a layer that doesn’t depend on Claude?

Every setting above relies on Claude getting the task right. OnlyRead doesn’t ask Claude anything. It sits in your browser, checks what each of Claude’s clicks would send to the site and refuses the ones that would change something, whatever Claude decided. On the sites you guard, such as LinkedIn, X or Linear, reads and searches go through, and a request that would send, post, delete or change a setting is refused before the app’s server sees it. So you can give Claude a reading job on LinkedIn and get on with your day, without watching every click.

What Claude believed it was clicking doesn’t matter, and neither does its mode or whether you approved the site. If Claude sends a LinkedIn reply it meant only to draft, or clicks Connect on a profile it meant only to read, the request is what gets checked, and it stops. Keep Claude’s safeguards on as well. The two layers catch different mistakes, and there’s no reason to switch Claude to Skip all approvals to use OnlyRead.

  1. YouChoose the sites Claude may use and the ones OnlyRead guards
  2. ClaudeAsks you to approve the websites it will use
  3. OnlyReadRefuses requests that would change something
OnlyRead covers only the sites you guard, in the Chrome where it’s installed.

Set it up for Claude in Chrome

  1. Install it in the Chrome you use with Claude, from the Chrome Web Store. Claude in Chrome runs only in Google Chrome. On a managed work browser, ask IT first.
  2. Connect a checking service. OnlyRead uses it for the few requests its built-in rules can’t decide. Sign in to OnlyRead Pro, or paste your own key from OpenRouter or Vercel AI Gateway. Protection switches on once one is connected.
  3. Guard the sites Claude will read. Every site starts off. Switch on the apps in the task, such as LinkedIn, X or Google Calendar, in the settings or with Guard this site in the popup.
  4. Keep the default timing, or choose Always. By default OnlyRead guards a tab once it notices an AI agent taking control of it, which is how Claude in Chrome works. Always, under Security, guards from the moment a page loads, and Chrome then shows its debugging banner the whole time.

The general guide covers the rest of the setup and how read-only browser protection works.

Reading can look like writing

Searching LinkedIn isn’t only loading pages. Typing in the search box, filtering results and opening a conversation all send something to LinkedIn, and some apps send a search the same way they send a message. OnlyRead looks at what each one does, so the search goes through and the send doesn’t. The general guide shows why a search and a send can look alike.

How do you check that a read passes and a change is stopped?

Run one read and one harmless change on a guarded site, and note which layer stopped the change.

  1. Use something you own. LinkedIn has no private sandbox, so test on an app where nothing reaches anyone else. A throwaway issue in Linear or a post on a private test X account works. Don’t test on a live inbox: no message to a real person, no public post, no delete and no purchase.
  2. Set Claude to Manually approve, then open OnlyRead’s popup on the site. Protection should be on, and the site set to On.
  3. Ask Claude to read, for example: “Find my test issue in Linear and summarize it.” Approve the websites Claude lists. The answer should come back as usual.
  4. Ask for the small change, such as changing the priority of your test issue or liking your own test post. Approve again so the click goes ahead. If Claude stops to ask anyway, that’s Claude’s own check, so choose Allow this action. A change that got past Claude is what OnlyRead is for.
  5. Look for OnlyRead’s notice on the page, naming the stopped change. Details shows why.
  6. Reload and check that nothing changed, then find the decision under Activity in OnlyRead’s settings.

If Claude declines on its own before it clicks, that’s Claude’s safeguard, not OnlyRead’s. Only the notice and an Activity entry show the request reached OnlyRead and was refused. Teams writing up a pilot should note the Chrome, Claude and OnlyRead versions, the panel and mode, and the app before and after, in test accounts whose messages can’t reach anyone.

What can the two layers still not guarantee?

Neither layer is a guarantee, and OnlyRead has limits of its own:

  • OnlyRead checks only guarded sites, in the Chrome where it’s installed. Cowork’s built-in browser in the Claude app is a separate browser, and connectors or APIs that reach your apps outside the browser aren’t seen.
  • With the default timing, what a page sends before Claude’s first action isn’t checked, and a live connection opened before then stays open until you reload. Always mode closes that gap.
  • Requests its built-in rules can’t settle go to the checking service, an AI model, and it can misjudge one. It runs apart from the agent and answers one question, whether a request changes data. When it’s unsure, or the service can’t be reached, OnlyRead refuses, so a doubtful call costs a retry rather than a sent message.
  • Guarding is per site, so while Claude works in LinkedIn, your own messages and posts there are refused too.
  • Someone at the computer can turn the extension off in chrome://extensions, unless Chrome enterprise policy prevents it.
  • Claude still reads everything on the pages it opens, and screenshots of its tabs become part of the conversation. Anthropic advises against using Claude in Chrome on sensitive sites,4 and read-only doesn’t change that advice.

Is a draft a change?

Typing a reply without sending it passes. But some apps save drafts to their server as you type, so a draft Claude writes can show up in your Drafts folder. Read-only here means no send, post, delete, purchase or setting change.

Who sees request data?

Most requests are decided on your device. Only a request nothing local can classify is described to a model: on Free, to the OpenRouter or Vercel AI Gateway provider behind your own key, and on Pro, to OnlyRead’s service. That description can include personal data from the request. The privacy policy lists exactly what is sent.

Questions

Isn’t Manually approve enough?

No. Once you approve LinkedIn as a site Claude may use, it acts there without asking at each step, so a send, post or connection request it makes by mistake goes ahead unless Claude itself notices. OnlyRead checks the request itself, so on a guarded site that send is still refused.

Does Automatically approve mean Claude has no safety checks?

No. It screens each action, blocks the ones it judges unsafe and pauses for you when needed. Skip all approvals is the mode without those checks. Neither promises that every send or post will come to you first.

Does OnlyRead work with Claude Cowork?

Yes, when Claude Cowork works in your Chrome. OnlyRead guards the sites you switch on there. The Claude app’s own built-in browser is a separate browser, and connector actions that reach your apps outside the browser aren’t covered, so keep Claude’s approvals on for those.

What if I want Claude to write on one site?

Switch that site off in OnlyRead before the task, keep Manually approve on and watch what Claude does there. LinkedIn and your other guarded sites stay read-only.

Sources

  1. Messages API, LinkedIn docs on Microsoft Learn, retrieved Nov 16, 2026.
  2. Claude in Chrome permissions guide, Anthropic Help Center, retrieved Nov 17, 2026.
  3. Get started with Claude in Chrome, Anthropic Help Center, retrieved Nov 12, 2026.
  4. Use Claude in Chrome safely, Anthropic Help Center, retrieved Nov 17, 2026.