Internal mascot bot for non-technical staff — read-only MCP + Gmail-draft-only write

A named, branded bot sits in Telegram and WhatsApp for internal staff. Internally it is a Claude Code agent with a strict business-context prompt, exposed via MCP for custom tools (because Codex needs MCP to receive custom tools), with two hard rules: every MCP tool is read-only, except Gmail — and even there the only allowed write is creating a draft, never sending. The bot has an analyzed feedback loop where the operator reviews chat transcripts and refines instructions, skills, and tools.

Two underrated wins: (1) the bot runs on the founder's existing Claude subscription rather than per-seat enterprise tooling, so cost stays predictable; (2) by being branded as a "mascot" rather than "AI assistant", non-technical staff actually use it — and importantly stop pinging the founder. The feedback loop on transcripts is the multiplier: instructions improve weekly.

1. Pick a name, mascot image, and personality consistent with the company brand. Non-technical adoption hinges on this more than on capability.
2. Stand up a Claude Code agent with a tight business-context system prompt (what the company does, who works there, what they're allowed to ask about).
3. Expose custom tools via MCP — Codex specifically requires MCP for tool access, so the indirection is forced anyway.
4. Hard rule: every MCP tool is read-only against the CRM, inventory, internal docs. No write tools that affect business state.
5. The one exception: Gmail draft creation. The bot can compose a draft email but cannot send. The human reviews and sends.
6. Wire the bot into Telegram and WhatsApp. Telegram bot framework is straightforward; WhatsApp needs an unofficial-API bridge or a Business API account.
7. Set up a semi-automated feedback loop: weekly, an operator (or another agent) reads transcripts, identifies where the bot missed, and updates instructions, skills, or MCP tool descriptions accordingly.

Tokens come out of the founder's personal subscription budget. One operator reported the Opus-quality bar is non-negotiable for production use; cheaper models (Gemini via Vertex at half the price was tested) "made things multiple times worse and still expensive". Budget for one $200/mo seat per ~10-20 active staff, and expect to keep upgrading subscription tier as usage grows. Second gotcha: Telegram does not have native threading (unlike Slack), so multi-turn session management for many simultaneous staff users is harder than Slack equivalents.