Sanbox By Sanlabs

Guides

Agents for teams

Available on request

Trigger an AI agent from Slack, Microsoft Teams or WhatsApp

By the Sanbox team updated

Ask ChatGPT Ask Claude

Sanbox connects your agents to Slack, Microsoft Teams and WhatsApp. Someone mentions the agent or sends it a message, and the work runs on the agent's own Computer, with its files, memory, tools and limits, and the answer comes back in the same conversation. These triggers are enabled per organisation on request. On WhatsApp, agents work within Meta's WhatsApp Business terms, which limit general-purpose AI assistants, so use it for business-specific workflows. Today, a Hermes Computer takes work by email and Telegram, and any system that can run a command can start work with the Sanbox CLI.

What do you get?

  • Agents where people already talk. A colleague asks in a channel or a direct message, and the agent answers there. Nobody needs to learn a new app to ask a question.
  • The right agent for each conversation. A sales channel reaches the sales team's agent, and an engineering channel reaches one with access to the code. Each agent keeps its own template, network policy and secret proxies.
  • Work that keeps going. The agent runs on a persistent Computer, so a long task continues after the message is sent, and files it creates stay on its disk.
  • The same model, your key. Every message is answered by the model provider you configured for that agent.

How does it fit with what you already run?

Triggers are another way in, not another kind of agent. The Computers, templates and controls are the ones you already use from the console and the CLI:

  • Network policy still decides which hosts the agent can reach.
  • Secret proxies still keep company API keys out of the Computer.
  • Stop and resume still work: a stopped agent doesn't answer until you resume it.
  • EU data centres still hold the Computers, their disks and run records. Messages in Slack, Teams or WhatsApp stay with those services under your own agreements with them.

Enable it for your organisation

Enable it for your organisation and tell us which of the three you use, how many people and channels should reach an agent, and what the agents should be allowed to do. We'll set it up with you.

What works today

Email and Telegram through Hermes. A Hermes Computer runs Nous Research's Hermes Agent, which keeps memory across sessions and learns reusable skills. Give it an address from your connected AgentMail account and a Telegram bot when you start it, and people reach it from any mail client or from Telegram. Only the Telegram user IDs you list can message the bot. The full recipe is How do I run Hermes Agent in the cloud with memory, skills, an inbox and Telegram?

Your own bot or workflow calling the CLI. If you already run a chat bot or an automation tool on a server, it can start an agent task with the Sanbox CLI and an organisation API key, wait for the result, and post it back. The pattern is the same as calling Sanbox from n8n: see Call an agent computer from n8n, Make or Zapier.

What to decide before you switch it on

  • Who can reach which agent. An agent in a busy channel gets questions from many people. Give it only the tools and data that everyone in that channel may use.
  • Which hosts it may call. Keep each agent's network policy to the APIs it needs.
  • What it costs. Each message is a model call on your provider account. Set limits there.

Sources