Run an agent as a real Telegram user account (gramjs / MTProto)

Connect an agent to Telegram as a regular user account (not a bot) using gramjs over MTProto. The agent gets a real phone number, joins groups as a member, sees full history, and answers in-line. For outbound messages to people the operator hasn't pre-authorized, the agent drafts and a separate bot posts the draft for human approval before sending.

Bots are second-class citizens on Telegram: they can't read group history before being added, can't be silent observers, and are visibly tagged. A user-account bridge makes the agent indistinguishable from a human team member in a work chat. The reported use case: an agent answers in shared work chats directly; for any new outbound conversation, drafts go through a side bot so the founder reviews before send.

1. Register a new app at my.telegram.org as a regular user app — get API_ID and API_HASH.
2. Provision a fresh phone number (a service like telnyx.com works) for the agent's account.
3. Auth the account once with gramjs and persist the resulting StringSession in a TG_SESSION env var.
4. Wire the bridge: gramjs listens via client.addEventHandler for NewMessage, pipes the message into your agent harness (chat-completions API), and posts the reply back via the same gramjs session.
5. For outbound to people not yet in the trusted whitelist: route through a separate Telegram bot that posts the draft + an inline approval button.
6. Buy Telegram Premium for the account — it raises rate limits and reduces flag risk in the first weeks.
7. In the first day or two, keep volume low and vary message phrasing to avoid spam-pattern detection.

• User-account access exposes the agent to a much larger prompt-injection surface: any message in any group it joins becomes potential input. Pair with a researcher/executor split (see B-058 / defense-in-depth pattern) — the gramjs listener should not have shell or write access; route output through a sanitizing gateway.