Claude Desktop's approval-policy deadlock: MCP tool calls need approval but policy is set to 'never'
danshipper · x · 2026-10-20
- Dan Shipper (Every CEO) says he spent an hour fighting security policies in Claude Desktop and the ChatGPT app instead of working, calling it the arrival of a "nanny state era for AI."
- Screenshot shows the exact deadlock: an MCP tool call requires approval, but the chat's approval policy is set to never — so the agent needs approval it can never request, and doesn't know how the user can change the policy. Tests kept failing with the same error.
- A concrete case of the friction between AI safety guardrails and usability.
Related event: Every CEO Slams Annoying Safety Policies in Claude Desktop and ChatGPT(2 posts)→
More from Fun
- This Concert Would Cost Millions to Stage — Made Entirely With AI — QwaZz · 2026-10-23
- Letting Claude agents chat freely ends in a deadlock traffic jam, dev warns — rickasaurus · 2026-10-23
- A Midjourney --sref code for crayon-style "cats in sunglasses" illustrations — sergeantsref · 2026-10-23
- Viral AI meme: bedraggled Barbie sips a drink in front of riot flames — McDonaghMatthew · 2026-10-23
- Nat Eliason: One call with Matt Shummer sent our token spend up 10x — nateliason · 2026-10-23
- 'Everyone in SF has AI psychosis': insiders urge stepping out of the AI bubble — TarlonKhoubyari · 2026-10-23