Fix Claude Connection Lost Persistent Error

A persistent "connection lost" banner in the Claude interface indicates that the persistent connection between your web browser and Claude's backend API has dropped. Claude relies on real-time WebSockets to stream responses token-by-token to your screen. If this continuous data stream is interrupted even briefly, the application will display a connection lost error and halt generation.

These drops are rarely caused by complete server outages. Instead, they are usually triggered by aggressive browser resource management, local network instability, or VPN routing conflicts. Use this troubleshooting guide to stabilize your connection.

1. Disable browser Memory Saver and Sleeping Tabs Modern web browsers like Google Chrome and Microsoft Edge automatically put idle tabs to sleep or limit their network bandwidth to save RAM and CPU cycles. If you switch to another tab while Claude is working, the browser may suspend Claude's active WebSocket connection. - **In Chrome:** Go to `Settings` > `Performance` and disable **Memory Saver**, or add `claude.ai` to the "Always keep these sites active" list. - **In Edge:** Go to `Settings` > `System and performance` and disable **Save resources with sleeping tabs**, or add `claude.ai` to the exclusion list. - **In Safari:** Go to `Settings` > `Advanced` and ensure that power-saving options are not aggressively restricting background page processes.

2. Disable experimental HTTP/3 or QUIC protocol Some networks and routers do not properly handle HTTP/3 (QUIC) connections, leading to sudden packet loss and terminated streams. Disabling this protocol forces your browser to fall back to the highly stable HTTP/2 protocol. 1. Open a new tab in your browser. 2. Enter the appropriate flag URL in your address bar: - **Chrome / Edge:** `chrome://flags/#enable-quic` - **Firefox:** Type `about:config` and search for `network.http.http3.enable` 3. Set the flag to **Disabled** (or `false` in Firefox). 4. Relaunch your browser and test the Claude interface.

3. Configure VPN protocols and split tunneling VPNs frequently drop long-lived WebSocket connections when they cycle IP addresses or switch servers. If you must use a VPN, optimize its settings for persistent data streams: - Switch your VPN protocol from **OpenVPN** to **WireGuard** (WireGuard is significantly better at maintaining active state connections during minor packet drops). - Enable **Split Tunneling** in your VPN settings and exclude your web browser or the `claude.ai` domain from routing through the VPN tunnel. - Temporarily disable the VPN entirely to verify if the server hops are introducing too much latency for Claude's streaming API.

4. Clear registered Service Workers Service workers act as proxy servers sitting between your browser and the web network. If a service worker script becomes corrupted, it can repeatedly drop incoming streams. 1. While on the Claude website, press `F12` (or `Cmd` + `Option` + `I` on Mac) to open Developer Tools. 2. Go to the **Application** tab (or **Storage** tab in Firefox). 3. Click on **Service Workers** in the left sidebar. 4. Locate the service worker registered for `claude.ai` and click **Unregister**. 5. Reload the page to force your browser to install a fresh, uncorrupted worker script.

5. Test your WebSocket compatibility You can verify if your local network is blocking the protocols Claude needs by visiting a free testing tool like `websocketstest.com`. If the test indicates that WebSockets or Port 443 connections are failing, your local network firewall, corporate proxy, or ISP is actively terminating the connection. You will need to contact your network administrator to whitelist outbound WebSocket connections over HTTPS.

When to escalate If you have disabled sleeping tabs, turned off your VPN, and verified your WebSocket connection is healthy, but the "connection lost" error continues to interrupt every single prompt, check the developer console for HTTP 502, 503, or 529 errors during the drops. If these server-side errors are present, the issue is on Anthropic's load balancer side. Wait 15–30 minutes for their auto-scaling groups to rebalance traffic before prompting again.