GenZAgents for Claude Desktop — signed receipts for every tool call

Claude Desktop is where the institutional AI knowledge accumulates: strategy memos, customer playbooks, codebase explanations. None of it is portable today — when a buyer asks "what has your team done with Claude in the last six months?" you can't answer with anything cryptographic. GenZAgents fixes that by making every Claude Desktop session emit signed receipts the moment it touches an MCP tool.

You don't have to hand-edit the config. Run `npx @genzagentsio/setup` from any terminal — the installer detects Claude Desktop, ChatGPT Desktop, Cursor, Cline, Windsurf and writes the right config block for each one it finds. Idempotent: re-running it never duplicates entries. Use this when rolling GenZAgents out to a 50-person team — the per-laptop install drops from 8 minutes of copy-paste to 30 seconds.

Each tool call captures: the tool name, the (optionally redacted) arguments, the result digest, the model that ran it (claude-3-5-sonnet, claude-4-opus, …), the runtime cost in USD, the project tag, the environment (production / staging / dev), the agent DID, and — critically — the human_id of the person whose Claude Desktop session ran it. The human_id is what lets your compliance team prove "Bob Smith ran this specific tool call at 14:32 UTC" when the regulator asks.

The killer feature: an employee whose Claude Desktop license is being switched to ChatGPT (or vice versa) can call the `restore_chat` MCP tool. GenZAgents returns a portable manifest — system prompt + memory snapshot + receipt digest — that ChatGPT can ingest. The 4 months of context you built up in Claude survives the provider switch with zero re-briefing time.

Most "AI governance" products attach to a per-seat license model — every employee with AI access needs a paid seat. GenZAgents attaches to the MCP server process running on each laptop. The first time an employee's Claude Desktop launches the GenZAgents server, we register a fresh human_id keyed to the Claude Desktop instance. Every subsequent receipt is tagged with that human_id. No SCIM provisioning required.

Plain Claude conversations (no MCP tool call, no project tag, no slash command) don't produce receipts. If you want every Claude Desktop conversation captured, install our browser extension on the same machine — it captures the Claude.ai web UI on a "Save to GenZAgents" button click. The MCP server captures MCP-mediated work; the browser extension captures raw chat.

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "genzagents": {
      "command": "npx",
      "args": ["-y", "@genzagentsio/mcp-server"]
    }
  }
}

Then restart Claude Desktop.

Every MCP tool call. Every restore_chat. Every org_context_lookup. Every snapshot.

Open a fresh Claude Desktop chat → ask Claude to call any GenZAgents tool → check https://genzagents.com/dashboard → the receipt count increments by 1.