Platform: Perplexity Pro, Desktop Web (Chrome), JP region
Date/Time: 2026-05-08 ~14:15 JST
Connector: mcp.workjournal.pro (remote MCP, OAuth-authenticated)
What Happened
Prompting “try listing my journals from workjournal” triggered a tool call that hung for ~2 min 6 s, then collapsed to a red “Sorry, something went wrong” banner. Re-adding the connector multiple times did not help.
Browser Console Errors
1. CORS failure: The restricted-feature-loader JS chunk (from pplx-next-static-public.perplexity.ai) was redirected through perplexity-ai.cloudflareaccess.com. The OPTIONS preflight returned no Access-Control-Allow-Origin header, so the script never loaded.
2. Stream failure: PplxStreamError type=STREAM_FAILED_PLACEHOLDER_ERROR, errorCode=unknown
Request IDs
- Request ID:
f0d26f8e-958a-4efe-bd79-825b46bc21a9 - Backend UUID:
22c0ef96-7f17-48f6-9081-6fb350879c90
MCP Server Verified Healthy (not the issue)
- discovery: 200 in 274ms
- unauthenticated POST: 401 in 486ms
- openid-config: 200 in 683ms
- JWKS: 200 in 203ms
- Same connector works on Grok, ChatGPT, Claude Desktop, and workjournal CLI
Hypothesis
The restricted-feature-loader chunk is inadvertently gated behind Perplexity’s Cloudflare Access policy. Browsers never send cookies with OPTIONS preflights, so the chunk fails to load silently, breaking MCP stream rendering even when the backend responds successfully. After ~2 min the pplx-stream layer times out and shows the generic error. This aligns with a Warn-level Perplexity Website+API incident recorded by StatusGator on May 8, 2026.
Research thread with full details
https://www.perplexity.ai/search/when-you-read-this-from-a-work-pCxX._nyQfuMJ1Y7yA7o_w