One OpenClaw Gateway Multiple Isolated AI Assistants (One Telegram Bot Per Worker)

#openclaw #agents #telegram

tldr;

Give every team member their own private AI employee on a single machine. Full isolation, zero extra servers, Telegram-only setup in minutes.

Abstract

Tired of spinning up separate servers or cloud instances for every AI assistant in your company?

OpenClaw lets you run dozens of completely separate, fully isolated agents on one single machine — and Telegram makes the UX feel magical.

Each worker gets their own dedicated bot (@AliceSalesClawBot, @BobSupportClawBot, etc.). They chat privately with “their” AI, have their own memory, files, tools, and personality — and they never see anyone else’s stuff.

Here’s exactly how I set it up.

Why this works so well

One Gateway process handles everything
Each agent gets its own workspace folder (~/.openclaw/agents/alice-sales/)
Separate SOUL.md, AGENTS.md, history, allowed tools, LLM keys, everything
Telegram routing is dead simple and rock-solid

From the worker’s perspective it’s just “my personal AI in Telegram”. They have no idea it’s all running on the same Mac Mini / VPS.

Step 1: Create one agent per worker

openclaw agents add alice-sales
openclaw agents add bob-support
openclaw agents add carol-finance
# ... one for every person

Step 2: Create a Telegram bot for each agent

Talk to @botfather
/newbot → give it a clean name
Copy the token

Do this once per worker.

Step 3: Configure everything in one file

Run openclaw configure or edit ~/.openclaw/openclaw.json directly.
Here’s a ready-to-paste example for a 3-person team:

{
  "agents": {
    "list": [
      { "id": "alice-sales", "workspace": "~/.openclaw/agents/alice-sales" },
      { "id": "bob-support", "workspace": "~/.openclaw/agents/bob-support" },
      { "id": "carol-finance", "workspace": "~/.openclaw/agents/carol-finance" }
    ]
  },
  "bindings": [
    { "agentId": "alice-sales", "match": { "channel": "telegram", "accountId": "alice" } },
    { "agentId": "bob-support", "match": { "channel": "telegram", "accountId": "bob" } },
    { "agentId": "carol-finance", "match": { "channel": "telegram", "accountId": "carol" } }
  ],
  "channels": {
    "telegram": {
      "enabled": true,
      "accounts": {
        "alice": {
          "botToken": "7123456789:AAFxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
          "dmPolicy": "pairing"
        },
        "bob": {
          "botToken": "7987654321:AAFyyyyyyyyyyyyyyyyyyyyyyyyyyyyy",
          "dmPolicy": "pairing"
        },
        "carol": {
          "botToken": "7843210987:AAFzzzzzzzzzzzzzzzzzzzzzzzzzzzzz",
          "dmPolicy": "pairing"
        }
      }
    }
  }
}

Security tip (do this immediately):

After the worker sends their first message, change dmPolicy to "allowlist" and add their Telegram numeric ID:

"allowFrom": ["tg:1234567890"]

(Get the ID by having them message @userinfobot)

Step 4: Restart & test

openclaw gateway restart
openclaw agents list --bindings

Send each worker their bot link. Done.

What you get automatically

Full isolation (memory, files, tools, history)
Different models or permissions per person (Claude for leadership, cheaper model for interns)
Agents can still talk to each other internally when needed
Works on anything: old Mac Mini, cheap VPS, even a Raspberry Pi with some tweaks

Official docs I used

Security hardening

If you want tighter control over your Openclaw, consider installing via my open source Openclaw hardening kit: https://github.com/NinoSkopac/openclaw-secure-kit

It is a turnkey solution - secure-by-default OpenClaw, with a verifiable security report.