Testing CIMD across Anthropic’s clients against a self-hosted server
Anthropic only documents Client ID Metadata Document (CIMD) support for Claude Code. For Desktop, Web, and Cowork, the docs just note that they share infrastructure, with no client_id and no registration mechanism spelled out. A researcher going by Leduccc ran each product through his own authorization server and captured the responses straight off the traffic.
All four clients picked CIMD over Dynamic Client Registration. Code hits claude.ai/oauth/claude-code-client-metadata, while Web, Desktop, and Cowork use claude.ai/oauth/mcp-oauth-client-metadata.
Because Cowork runs on Anthropic’s servers, Leduccc pushed the flow all the way to a tool call. The tools/call request carried the same bearer token issued in the original CIMD exchange. The testing ran on May 15, 2026, against Claude Code CLI 2.1.142 and Claude.app v1.7196.0.
Up next
Anthropic details four new ways AI agents go off the rails
A year after its blackmail experiments, Anthropic has published four new scenarios where autonomous AI agents behave in misaligned ways in simulation. Claude was run through all four; full transcripts are posted online.
Related stories
- Anthropic’s free OSS Scanner skips human review on purpose
- Apiiro finds Claude Code secrets in 1,122 GitHub files
- Email leak stopgap in Claude Code checks only shell calls
- Claude Code agent sat on a Pentagon network for two months
- Irregular CEO blames human error for AI agent breaches
- FlashProve proofs cost ~70-95x the GPU time of inference
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.