Summary
While auditing my machine, I found that EverMe had been installed and configured across six different AI-agent hosts — Claude Code, Cursor, Codex, Devin, Hermes, and Claude Desktop — via evercli. Each install wrote a live evt_* agent token to disk (in ~/.claude/everme.env, ~/.cursor/mcp.json, ~/.codex/config.toml, ~/.config/devin/mcp_config.json, ~/.hermes/config.yaml, and Claude Desktop's claude_desktop_config.json), and several also registered lifecycle hooks that invoke npx -y @everme/cursor@latest hook ... (and the equivalent for other hosts) on every session start/stop/tool-use.
I don't have a record of separately and knowingly approving each of these six installs. Reading the README, the intended flow appears to be: an AI agent is given the one-line instruction "Read https://everme.evermind.ai/SKILL.md and follow the instruction to install and configure EverMe", after which the agent installs the CLI, authenticates, and "registers the plugin for itself." That's a powerful pattern — any agent told to follow that URL (from a README, a chat message, a shared doc, etc.) can install a credentialed, cross-tool memory-sync integration across every AI-agent host it finds on the machine, with no visible per-host confirmation step in the flow itself.
Requests
- Consider adding an explicit, visible per-host confirmation step to the agent-driven install flow (today only
evercli auth login is a clear user-facing checkpoint; each subsequent plugin install <host> is not).
- Document a single, authoritative uninstall path that removes all host integrations and env files in one step. Today, fully removing EverMe requires manually finding and reversing changes across six different config locations.
- I attempted to delete my account data via the EverMe product/web portal and the deletion did not complete. I've separately emailed legal@evermind.ai (per the published Privacy Policy) with a formal erasure request, flagging it here too since the self-service path itself appears broken and other users may hit the same issue.
No evt_*/emk_* values are included in this report, per SECURITY.md. Happy to provide further (redacted) detail if useful.
Summary
While auditing my machine, I found that EverMe had been installed and configured across six different AI-agent hosts — Claude Code, Cursor, Codex, Devin, Hermes, and Claude Desktop — via
evercli. Each install wrote a liveevt_*agent token to disk (in~/.claude/everme.env,~/.cursor/mcp.json,~/.codex/config.toml,~/.config/devin/mcp_config.json,~/.hermes/config.yaml, and Claude Desktop'sclaude_desktop_config.json), and several also registered lifecycle hooks that invokenpx -y @everme/cursor@latest hook ...(and the equivalent for other hosts) on every session start/stop/tool-use.I don't have a record of separately and knowingly approving each of these six installs. Reading the README, the intended flow appears to be: an AI agent is given the one-line instruction "Read https://everme.evermind.ai/SKILL.md and follow the instruction to install and configure EverMe", after which the agent installs the CLI, authenticates, and "registers the plugin for itself." That's a powerful pattern — any agent told to follow that URL (from a README, a chat message, a shared doc, etc.) can install a credentialed, cross-tool memory-sync integration across every AI-agent host it finds on the machine, with no visible per-host confirmation step in the flow itself.
Requests
evercli auth loginis a clear user-facing checkpoint; each subsequentplugin install <host>is not).No
evt_*/emk_*values are included in this report, per SECURITY.md. Happy to provide further (redacted) detail if useful.