A runtime gate for coding agents: an MCP server that blocks the tool calls that wreck a repo (force-push main, rm -rf, secret exfil, CI wipe) and lets real work through. Machine-checked git-branch core. Tested on 3,790 real commands.
Privacy-preserving feedback co-author for Claude: redact secrets/PII/IP from your session, then ship clean feedback through Anthropic's real /feedback intake — non-destructively. In-session co-author + MCP toolbox across CLI, API, claude.ai, Cowork, IDE. For the Anthropic Feedback Loops role.
MCP server that lets Claude Code generate & edit images through your own ChatGPT and Gemini web accounts — no image API keys. Playwright-driven persistent browser sessions; you handle login/CAPTCHA yourself; Claude reviews the images.