[BUG] indows 11 Pro: Cowork, Connectors, and Claude Code all broken after latest auto-update — ERR_CONNECTION_RESET + OAuthError

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/46818
2. https://github.com/anthropics/claude-code/issues/46123
3. https://github.com/anthropics/claude-code/issues/39225

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

Same repro on Windows 11 + latest Claude Desktop auto-update.

• Connector: custom remote MCP server (OAuth 2.1 + PKCE, stateless HMAC tokens, deployed as a Cloudflare Worker).
• UI error: "Couldn't reach the MCP server" — reference ofid_6520722aaff2b914.
• Console: TypeError: Failed to fetch originating from assets-proxy.anthropic.com/claude-ai/v2/... (same asset-proxy fetch failure you listed), plus the unrelated No handler registered for ...BuddyBleTransport_$_reportState noise.

Server is confirmed healthy and reachable:

• Direct curl against /health, /.well-known/oauth-protected-resource, and POST /mcp return 200 / 200 / 401 (expected OAuth challenge) from the same machine that's running the broken Claude Desktop.
• wrangler tail on the Worker sees zero requests from Claude Desktop during reconnect attempts, while the same wrangler tail does see a manual curl from the same machine.

So the request never reaches Cloudflare's edge — the failure is in the Desktop → Anthropic backend hop before the MCP server is contacted at all. This matches the pattern reported in anthropics/claude-ai-mcp#143.

Only change on this machine between "working" yesterday evening and "broken" this morning: the Claude Desktop auto-update (and a concurrent Windows 11 cumulative update). Pre-existing HTTP proxy / AV / DNS setup did not change, and disabling the proxy made no difference.

(#47104) and ask Anthropic directly for the last known good installer URL. They host installers at predictable paths like https://storage.googleapis.com/osprey-downloads-c02f6a0d-347c-492b-a752-3e0651722e97/nsis-updater/Claude-Setup-x64.exe — but the versioned URLs aren't public.

Thanks @DeeBeeJason — that's a solid suggestion.

Confirming same regression on Windows 11 Pro, Max Plan (€89/month).
All three features broken after auto-update:

• Cowork: ERR_CONNECTION_RESET
• Connectors: "Could not load connectors directory"
• Claude Code in Desktop: OAuthError

I've since installed Claude Code CLI (v2.1.110) which works
perfectly from the same machine — confirming this is 100% a
Desktop-specific issue.

I've emailed support@anthropic.com twice now with zero response.
I'll ask them for the last known good installer URL as well.

@anthropics — can you please:

1. Provide the download URL for the last stable Windows installer

before this broken release?

1. Give us an estimated timeline for the fix?
2. Acknowledge this is a known regression?

5+ days of broken premium features on a $100/month plan with no
support response is not acceptable.

Using windows 11, was working normally, once i installed Version 1.3036.0 (cf521c), cowork stopped working... Commenting just to keep track if they give us any response.

Hey i just did a workaround and it worked, hope it helps: https://github.com/anthropics/claude-code/issues/49917#issuecomment-4268321062

same issue please fix

I'm experiencing the same issue. It works fine in other sessions, but only in the affected session do I get the error: "Error invoking remote method ‘$eipc_message$_4ab9ae55-583a-4867-90be-23b2daff8899_$_claude.web_$_LocalSessions_$_sendMessage’: OAuthError: Unable to start session. Check your internet connection and try again."
Could it be because I've hit the rate limit? I'm using additional usage, though...

Translated with DeepL.com (free version)

Adding a macOS + Starlink data point — same triple-failure pattern, still occurring as of May 12, 2026.
Environment:

macOS, Mac Mini
Claude Desktop, latest auto-updated build
ISP: Starlink
Location: Australia
Paid plan

Errors (both fire together, every session):

"Server is busy. Retrying in 1s (attempt 3 of 10)" on any Cowork action
Popup: "Could not load connectors directory"

Duration: Multiple days, persistent.
Critical diagnostic: Works perfectly on iPhone cellular hotspot, fails on Starlink. Confirmed via curl -v that api.anthropic.com and assets-proxy.anthropic.com are reachable from Starlink (clean TLSv1.3 handshake, same resolved IP 160.79.104.10 as hotspot). Underlying network is fine — only the Claude desktop app fails. So this is an app-level interaction with something Starlink-specific, not endpoint reachability.
Already tried:

Cmd+Q + relaunch (many times)
Sign out / sign in
Cleared Cache, Code Cache, GPUCache, Session Storage, Local Storage
Latest desktop build confirmed
Cloudflare DNS on Mac and Starlink router
Starlink router reboot
macOS DNS cache flush
Disabled IPv6
Lowered MTU to 1420

Status page: Cowork shows Operational.
This isn't Windows-specific (original report was Win 11) — same symptoms on macOS. Suspect MTU/path-related interaction with Starlink's network stack for long-lived WebSocket/streaming connections, but standard MTU lowering didn't resolve it. Happy to provide network logs or request IDs.
cc @anthropics team

Update: same errors now occurring on iPhone hotspot as well as Starlink, so the issue is not network-related. Cowork tool calls to the Chrome extension (tabs_context_mcp) fail with 'Server is busy' even with Chrome open and extension enabled.

Another Windows 11 user blocked by this — 8 days and counting.

What's broken

Settings → Desktop app → General → Browser Use shows "No browsers connected" with the message "No Chrome instances are connected. Open Chrome with the Claude extension and sign in."

The Claude in Chrome extension is installed, enabled, and signed into the same Claude account as the desktop app. It works fine inside Chrome on its own. But Cowork can't see it. Clicking Recheck does nothing.

Screenshot attached.

Timeline

• Tuesday, May 5, 2026 (daytime): Cowork's Chrome connector working normally.
• Tuesday, May 5, 2026 (evening): Stopped connecting. Has not reconnected since.
• May 13, 2026: Still broken. 8 days down.

Environment

• Windows 11
• Latest Claude Desktop (auto-updates on)
• Latest Google Chrome
• Paying customer

What I've tried (none worked)

• 5 full uninstalls and reinstalls of Claude Desktop
• Reinstalled the Claude in Chrome extension
• Signed out and back into both the extension and the desktop app
• Confirmed same account on both sides
• Restarted Chrome
• Quit Claude Desktop from the system tray and reopened
• Toggled the Claude in Chrome connector off and on
• Clicked Recheck repeatedly
• Rebooted the computer

Support experience

Multiple tickets submitted at support.claude.com over the past 8 days. Every response has been from the AI agent saying it can't help and that a human will follow up. No human has followed up.

Ask

This looks like the same regression already reported in #58172 and #48806. Can someone from Anthropic confirm whether a fix is coming, share an ETA, or give a tested way to downgrade Claude Desktop to a working version? Paying users need a workaround in the meantime.

<img width="1920" height="1200" alt="Image" src="https://github.com/user-attachments/assets/73f200b2-96db-4dd1-b810-ccb3250c76ac" />

"Cowork extension installs and authenticates successfully but Claude Desktop never detects browser (Chrome OR Edge, fresh Windows installs, 5+ retries) — appears to be Windows native messaging or VM-level failure, not browser-specific"

I have the same issue. I can't use Claude Code in Desktop or Powershell.
I get these messages:

In Desktop:
Something went wrong
Try sending your message again. If it keeps happening, share feedback so we can investigate.
OAuthError: Unable to start session. Check your internet connection and try again.

In PowerShell:
Unable to connect to Anthropic services
Failed to connect to api.anthropic.com: UNABLE_TO_VERIFY_LEAF_SIGNATURE
SSL certificate error (UNABLE_TO_VERIFY_LEAF_SIGNATURE). If you are behind a
corporate proxy or TLS-intercepting firewall, set NODE_EXTRA_CA_CERTS to your CA
bundle path, or ask IT to allowlist *.anthropic.com. Run /doctor for details.

Same regression on Windows 11, confirming with my data.

Environment

• OS: Windows 11
• Install type: MSIX (package: Claude_1.9659.2.0_x64__pzs8sxrjxfjjc)
• Claude Desktop version: 1.9659.2.0
• Plan: Max
• Internet: fine — regular Chat works normally

Symptom
Claude Code in Desktop fails to start a session:
OAuthError: Unable to start session. Check your internet connection and try again.
Started this morning (~May 30, 2026) immediately after a Claude Desktop auto-update.
Only the OAuth/session features are affected; plain Chat works.

Logs (main.log)
[warn] [oauth] network error on fetch, flushing connection pool and retrying: net::ERR_CONNECTION_RESET
[error] oauth authorize network error: net::ERR_CONNECTION_RESET
[info] [oauth] failed to obtain oauth token (network_error): net::ERR_CONNECTION_RESET
[error] Cannot start session: oauth failed (network_error): net::ERR_CONNECTION_RESET
[error] Session initialization failed: Unable to start session. Check your internet connection and try again.

Notes

• Regression: yes, worked before today's auto-update.
• MSIX install, so rolling back to a previous version isn't feasible (no published prior package; Store serves only latest) — no user-side workaround.
• Standalone Claude Code CLI (browser OAuth) is unaffected and works — issue is specific to the Desktop app's embedded OAuth.

Closing for now — inactive for too long. Please open a new issue if this is still relevant.