Your AI Agent Can't Read Anything You Log Into — Why, and Why Cookies Aren't the Fix
Your AI Agent Can't Read Anything You Log Into — Why, and Why Cookies Aren't the Fix
You ask Claude Code to summarize the decisions in a Confluence page your team wrote last week. You paste a Notion link into Cursor and ask what the spec says. You point an agent at a Jira issue and ask for the reproduction steps.
It fails every time, and the failure is strange: the agent does not say "access denied." It often says something vaguely plausible, or reports an empty page, or describes a login form. Nothing in the error message explains what went wrong — because from the fetcher's point of view, nothing did.
Short answer
The page you are looking at exists only for an authenticated session, and your agent does not have one. Its fetch succeeds — HTTP 200, no error — and returns the page a signed-out stranger would see: a login screen, a redirect stub, or an empty application shell. This is a different problem from anti-bot blocking, and it has a different fix. Exporting your cookies into a headless browser is the obvious idea and the wrong one: it converts a reading problem into a credentials problem. The approach that works is to leave the session where it is and have the agent read through the browser that already holds it.
Three failure modes that look identical and aren't
Almost everything written about agents and inaccessible web content collapses into one story: "sites block bots." That story is true about Reddit and misleading about your Jira instance, and the conflation is why people reach for proxy rotation to solve a problem proxies cannot touch.
1. Anti-bot blocking. The site actively refuses your request. Datacenter IP ranges are challenged or refused, TLS fingerprints are inspected, bot-detection layers sit in front of the origin. Reddit is the canonical case. The site does not want a crawler and says so. We covered this one separately in letting your agent read the sites that block crawlers.
2. Client-side rendering. The site serves a shell and fills it with JavaScript afterward. A plain fetch gets a 200 and a skeleton. Nobody is blocking anything; the content simply has not been built yet at the moment the fetcher reads.
3. Authentication. The site is entirely willing to serve this page. It will serve it to you, instantly, because you are signed in. It will not serve it to an anonymous HTTP request, because that request is not anyone.
The third is the one nobody writes about, and for most knowledge workers it is the one that actually bites. You are not trying to scrape a hostile forum. You are trying to get an agent to read the wiki your own team maintains.
Note that the fix for each is different. For (1) you need a request that does not look like a crawler. For (2) you need a real rendering engine. For (3) you need a session — and that is the one you cannot obtain by changing vendors, because no vendor has your login.
The measurement: nine tools, nine HTTP 200s, almost no content
I fetched each of these server-side on 19 November 2026 with a current desktop Chrome user agent, followed redirects, then stripped tags and counted the remaining text.
| Tool | HTTP | Text chars | What came back |
|---|---|---|---|
| Atlassian admin | 200 | 14 | Administration |
| Linear (inbox) | 200 | 15 | Linear Loading… |
| Figma (files) | 200 | 20 | Login \| Figma |
| HubSpot | 200 | 24 | HubSpot \| Redirecting... |
| Asana | 200 | 307 | Login screen |
| Trello | 200 | 308 | Resource-loading notice |
| GitHub (settings) | 200 | 632 | Sign-in page |
| Slack (web client) | 200 | 906 | Marketing page + "Sign in" |
| Google Docs | 200 | 1,046 | Sign-in page |
Every single one is a success by the only signal most tooling checks. Not one contains the page.
This is the trap worth internalizing: a 200 with 14 characters of body text is a failure that reports itself as a success. If your pipeline logs status codes, it will look healthy while delivering nothing. Worse, if those 14 characters reach a language model, the model will sometimes write a confident paragraph about a page it never saw. An error would have been better.
The one outlier is instructive. LinkedIn returned HTTP 403 with Just a moment... Enable JavaScript and cookies to continue — a bot challenge. That is failure mode (1), not (3), on a site that happens to also have a login wall. Both walls are real; they are different walls.
Why cookie export is the wrong fix
Everyone arrives at the same idea within about ten minutes: export the cookies from Chrome, inject them into a headless browser, point the agent at that. It is a reasonable instinct. It is also how you turn a mild inconvenience into a security incident.
A session cookie is a bearer credential. It carries no scope. There is no read-only version, no per-document restriction, no "just this Notion page." Whoever holds a valid session cookie is you, with everything you can do — including changing settings and reading things you never intended to share. A credential that cannot be narrowed should not be copied casually.
Many enterprise sessions are bound, and replaying them breaks things. Modern SaaS increasingly ties sessions to device and network signals, and the industry is moving further that way — Device Bound Session Credentials bind a session to a hardware-held key specifically so a copied cookie is useless elsewhere. Even without that, a cookie exported from your laptop and replayed from a headless Chromium in a datacenter presents a different IP, a different fingerprint, and a different user agent. Sometimes it is rejected. Sometimes it works and trips your company's session-integrity alerting, which is a worse outcome: now you are explaining to your security team why your account appeared to be accessed from two countries. Security teams treat exactly this pattern — same session, new IP and fingerprint — as the signature of cookie theft.
Revocation is all-or-nothing. If an exported cookie ends up somewhere it should not be — a log file, a config committed by accident, a crash dump — the only remedy is invalidating the session server-side. That logs you out too. There is no "revoke the agent's copy," because there is no agent's copy; there is one session with two holders.
SSO makes it worse, not better. SAML and OIDC sessions are deliberately short-lived and refresh through a flow a headless browser cannot complete unattended. You end up re-exporting cookies by hand every few hours, which is both miserable and a standing invitation to automate the export — i.e. to write a script that harvests your own credentials on a timer.
None of this means cookie export never works. It means the cost is paid in credential exposure, and the thing you wanted was to read a wiki page.
What actually works: leave the session alone
The insight is small and it changes the shape of the problem. You do not need the agent to have your session. You need it to get the text of a page your session already renders.
Those are very different asks. The first requires moving a credential. The second requires moving a string.
Your Chrome is already signed in. It already passed MFA. It already holds whatever device-bound key the SaaS vendor issued. The page is already rendered on your machine, fully, correctly, with JavaScript run to completion. Agent Bridge lets an agent ask that browser to open a URL and hand back the rendered content as Markdown.
What crosses the boundary is the extracted text. What never crosses it is the cookie, the token, the password, or the session itself.
npm i -g web2md-mcp-server && web2md-mcp-setup
Then fully quit Chrome — Cmd+Q on macOS, tray icon on Windows — and reopen it. Chrome registers native-messaging hosts only on a cold start; closing the window is not enough. Your agent then has agent_convert for one URL and agent_batch_convert for up to 50. Setup details are in the Agent Bridge docs.
The trust model is worth stating plainly, because "an extension that reads your logged-in pages" deserves scrutiny rather than a shrug:
- The agent sends a URL and receives Markdown. It never receives credentials.
- Pages are opened in the browser profile you are already using. No second session exists.
- Nothing is proxied through a server — the extraction happens locally, on the already-rendered DOM.
- If you are not signed in to something, it stays unreadable. There is no mechanism here that could change that.
That last point is the honest boundary of the whole approach, and it is also the answer to the ethics question. This is not a bypass. A bypass gets you access you do not have. This reads a page you are already entitled to read, on a screen where it is already displayed. The nearest honest analogy is Reader Mode, or saving a page locally — delegating a read you could do yourself, just faster and in a format an agent can use.
Honest limits
- A real Chrome must be running, signed in, on your machine. Headless agents on servers get nothing from this. If your workload is unattended and server-side, this is the wrong tool and no amount of configuration will change that.
- Canvas applications stay unreadable. Figma's editor surface draws to a canvas, not to DOM text. A text extractor has nothing to take. The same applies to most design and whiteboard tools once you are inside the editing surface rather than a file list.
- It is sequential, around 6 seconds per URL. A real browser opens real tabs one at a time. Fifty URLs is roughly five minutes. A server-side crawler doing public pages in parallel is far faster — if your pages are public, use one.
- Coverage varies per tool. Notion, Confluence, Jira, Linear, and ordinary internal dashboards render as DOM text and extract cleanly. Test the specific tool you care about rather than trusting a general claim, including this one.
- PRO plan and the native messaging host are required for Agent Bridge specifically; single-page conversion in the extension is not.
When you should use something else
If your content is public and you need breadth — thousands of URLs, in parallel, on a schedule — a server-side crawler like Firecrawl or Jina Reader is genuinely the better instrument, and browser-based extraction cannot match its throughput. The browser route is not competing on speed. It competes on reach: the set of pages that exist only for an authenticated session, which no datacenter fetch can enter regardless of how fast it is.
And if you only need one page occasionally, you do not need any of this. Convert it in the extension and paste the Markdown.
Related reading
- Let your agent read the sites that block crawlers — the anti-bot half of the problem, with measured coverage
- Why AI can't access Reddit, X, and Substack — the platform-by-platform breakdown
- Let your AI read the pages it can't reach — the measured proof table, including Reddit, Skool, and Economist
- Agent Bridge documentation — setup, requirements, and the full tool reference
Related Articles
One Command Connects Web2MD to Claude Code, Cursor, Codex, and Claude Desktop
Let Your Agent Read the Sites That Block Crawlers: Reddit, HN, and More
Give Your AI Agent Reddit Access: Batch-Read Threads via MCP
Most Read
last 30 days- #1Can Claude Read Reddit? No — Here's the One-Click Fix (2026)
- #2HTML vs Markdown for LLMs: I Wasted 67% of My Tokens for a Year
- #3Reducing Token Waste in ChatGPT and Claude: 7 Techniques That Cut Costs 72%
- #4Obsidian Web Clipper Official Plugin 2026: Complete Guide + When You Need More
- #5Reddit JSON API vs Scraping: The Honest 2026 Comparison for Developers
Latest Articles
- 2026-11-20How to Feed Web Pages to Gemini 4 Neon as Markdown
- 2026-11-17How to Back Up Claude Chats to Markdown, Then Delete Them Safely
- 2026-11-16Can Claude Scan Websites? What It Can Read, What It Misses, and a Workaround
- 2026-11-15Como Transformar um Arquivo em Markdown: What Actually Works for DOCX, PDF, HTML, and Web Pages
- 2026-11-10ChatGPT to Markdown Chrome Extension: How I Export Chats Cleanly