Claude Not Working With a VPN? The Problem Was the Route, Not the Prompt

Claude accepted my login, displayed the chat window and then refused to answer. I was in an Istanbul airport hotel, trying to turn six interview transcripts into a board-ready risk summary before a client call. The prompt remained on Sending, then the page returned a generic network error. I blamed the browser, cleared its cache and signed in again. The same failure returned before the first sentence appeared.

The client meeting began in forty-one minutes.

The transcripts had already been cleaned of names and identifying details. I had read them once, highlighted the repeated complaints and drafted the outline myself.

What I needed from Claude was narrower: group the recurring risks, challenge my categories and help me compress sixty pages of notes into a structure I could verify before the call.

The hotel Wi-Fi was fast.

Email worked.

The client’s document portal worked.

Claude’s homepage worked.

The conversation did not.

My VPN also displayed Connected, which made the failure more confusing. The internet was present, the account was active and the page was open.

Only the work inside it had stopped.

The country was supported, but the route was not working

My first suspicion was regional access.

It was an easy explanation. I was traveling, the account had been created elsewhere and the VPN exit location had changed several times since breakfast.

But Türkiye was on Anthropic’s supported-location list. (Claude Help Center) I was not trying to make an unsupported country look supported.

Anthropic’s troubleshooting guidance also points users toward VPNs, firewalls and network restrictions when login or connection errors appear. (Claude Help Center)

So I disconnected the VPN.

Claude responded immediately.

I entered a short test prompt:

Summarize the difference between operational and reputational risk.

The answer began appearing at once.

That eliminated the prompt, browser and account as likely causes. Claude itself was available; the failure appeared only when my usual VPN route was active.

Turning the VPN off was useful as a diagnosis, but it was not an acceptable final setup. The client required me to keep work traffic protected on public hotel networks.

I needed Claude and the VPN to work together.

The established provider repeated the same failure in different countries

I had used the same major VPN provider for years.

It had a large support operation, a long public history and servers across every region I normally needed. Those were good reasons to trust it while traveling.

I selected Germany.

Claude loaded, but the prompt stayed on Sending.

I changed to the Netherlands.

The login page asked me to verify the session again. After I completed the check, the conversation opened and failed when I attached the first transcript.

I tried London.

That route accepted the upload.

Then the provider reconnected through another endpoint when the hotel Wi-Fi paused. Claude returned to the login screen, and the uploaded file disappeared from the conversation.

I had reached Claude three different ways without completing one useful exchange.

The server list now felt less like flexibility and more like a tray of nearly identical keys. Each one opened the front door. None kept the workspace available long enough to finish the job.

That changed the standard I was using. I did not need a server that made Claude appear for thirty seconds.

I needed one route that remained accepted from login through file upload and final response.

A shared exit address carries everyone’s reputation

A commercial VPN server can carry traffic for many users through the same public address.

One person may be reading email.

Another may be creating accounts.

Another may be sending automated requests.

The service receiving that traffic sees the shared exit point rather than each user’s reason for connecting.

Other Claude users have described the same practical split: one connection receives a network-block message, while changing the IP restores access. (Reddit)

That was enough to explain why random country switching was not solving the problem. The nearest server was not automatically the cleanest or most stable route for Claude.

What mattered was continuity.

Could I sign in once?

Could I upload all six files?

Could Claude finish the response?

Could the session remain intact while I verified the result?

So far, the established provider had not carried that sequence.

The phone hotspot changed the network, not the outcome

I switched the laptop from hotel Wi-Fi to my phone’s roaming hotspot.

The major VPN connected to Germany again.

Claude answered the first short prompt.

I attached the smallest transcript.

The upload reached 44 percent and paused.

The mobile signal inside the room moved between two bars and one. When the VPN reconnected, Claude asked me to refresh the page.

The transcript vanished.

I carried the laptop toward the window and restarted the upload.

This time, it reached 71 percent before the hotspot stalled.

The meeting began in twenty-six minutes.

The hotspot had proved that Claude could work briefly through the major provider. It had also made the real failure clearer: brief access was useless if the session disappeared halfway through the work.

I needed to upload all six transcripts, receive a structured response and remain connected while checking the output against my notes.

One successful test prompt did not complete any of that.

The browser extension protected the wrong workspace

I installed a free browser VPN extension with a nearby European route.

Claude opened.

The first transcript uploaded quickly.

The extension had one obvious appeal: it cost nothing and required almost no setup.

Then I opened the client’s desktop document application to compare Claude’s categories with my draft.

That application was outside the browser tunnel.

So were the email client, cloud-sync folder and meeting software.

The extension had protected one tab while leaving the rest of the work session on the direct hotel connection.

I could have copied material manually between apps, but that defeated the reason I was using a full VPN in the first place.

Then the extension changed routes.

Claude returned to the login page.

I closed it with nineteen minutes remaining.

By then, the requirement was precise: one full-device connection, accepted by Claude, that would stay in place through every upload and response.

The smaller app began with the situation

I had installed OnlydogVPN^↗ before the trip but had not made it my default.

The established provider had more locations, more reviews and a much longer public record. The smaller service’s shorter history was why I had treated it as a backup.

But I did not need another country list.

I needed a route for sensitive work on an unreliable public network.

I selected that situation in the app.

The connection established on the hotel Wi-Fi.

Claude opened without returning me to the login page.

I attached the first transcript.

The upload completed.

Then the second.

Then all six.

I entered the working prompt:

Group the recurring risks into no more than five categories. For each category, identify the evidence pattern, the likely business consequence and one question the board should ask. Do not invent numbers or claims that are absent from the transcripts.

Claude began responding.

The first category matched my notes.

The second combined two complaints I had separated.

The third identified a dependency I had underweighted: customers were not only frustrated by delays; they no longer trusted the status updates explaining those delays.

The full response finished without a network error.

I copied the structure into my draft and checked every category against the source transcripts.

Two points were too broad, so I removed them.

One board question was better than mine, so I kept it.

With nine minutes left, I uploaded the revised summary to the client portal.

The document status changed to:

Ready for review.

That completed the task.

The useful route stayed boring

The service uses an HTTP/3-based connection designed to remain stable on weak or changing networks.

The practical difference was already visible.

The major provider opened Claude on several routes but repeatedly lost the login, upload or conversation.

The browser extension protected only one tab and changed routes before the work was finished.

The backup kept one full-device session usable from the first transcript to the uploaded board memo.

I could not observe Claude’s internal network or abuse-filtering rules. I could compare what happened after each VPN connected.

For this problem, one route Claude continued to accept mattered more than the fastest ping or the largest country list.

The hotel Wi-Fi dropped during the call

The client joined the meeting.

I shared the summary and walked through the five risk categories.

The operations director challenged the third one.

“Are we sure customers distrust the updates, rather than simply disliking the delays?”

I reopened the transcript notes and found three examples that separated those ideas.

Then the hotel Wi-Fi disappeared.

The call froze.

My laptop moved to the phone hotspot.

The VPN recovered without sending Claude or the client portal back to their login pages.

The meeting returned.

I finished reading the examples, and the operations director said, “Keep that category.”

The main task had already succeeded. The network recovery prevented the live review from becoming a second failure.

That mattered because Claude was only one part of the workflow. The notes, meeting and client portal also had to remain available after the underlying connection changed.

The route did not merely produce one answer.

It kept the work around that answer intact.

The phone handled the final revision

After the call, the client asked me to change one heading before the document went to the board.

I had already closed the laptop and was walking toward the terminal.

The smaller app was installed on my phone but had not been configured.

I displayed a verification code on the laptop, entered it on the phone and connected without creating another conventional password-based account.

The client portal opened.

I changed Customer Communication Risk to Trust Erosion During Delays, saved the document and sent the final link.

The client replied:

Approved. Circulating now.

The verification code was not why Claude had worked.

It removed the smaller delay that appeared after the main job was complete, when one final edit arrived on a second device.

“Claude works” was too small a test

Before that morning, I had tested AI tools with one question:

Does the page answer?

That was not enough for real work.

The complete session required more:

Could I remain signed in?

Could I upload the files?

Could Claude finish the response?

Could I compare it with the source material?

Could I move the verified result into the client portal?

Could the session survive a change from Wi-Fi to mobile data?

The major VPN passed some of those steps.

The browser extension passed fewer.

The smaller backup carried all of them as one sequence.

That was why disabling the VPN helped diagnose the failure but did not solve it. It proved Claude was available. It did not create a protected working session I could keep.

The board memo settled the comparison

The established provider remained the larger and more familiar company. It offered more locations, more public reviews and years of operating history.

Its tested routes also turned one Claude session into repeated logins, missing uploads and unfinished prompts.

The free extension briefly opened the workspace but protected only the browser.

The smaller service had fewer regions and a shorter public record.

It was also the option that uploaded the transcripts, completed the analysis, kept the client portal connected and recovered when the hotel Wi-Fi failed.

I began the morning searching for a VPN that would make Claude load.

The board memo revealed the better standard: the useful route was the one Claude forgot to interrupt.