# Use Agent host with your AI

> Connect Claude, ChatGPT or Grok to Agent host through Geenius OS: it lists agents, talks to them, reads their digests and approvals queue and checks usage. A person decides.

Source: https://agenthost.geenius.io/docs/use-with-your-ai/

Agent host has no AI you must use to run it, but your own AI, Claude, ChatGPT or Grok, can work with your agents through Geenius OS: it lists them, talks to a deployed agent, reads a digest, looks at the approvals queue and checks usage. **A person decides** every approval.

> [!HONEST]
> The door can watch and talk, and it cannot decide. A risky action an agent wants to take, such as a webhook or a connector tool not marked read-only, comes back as pending approval, and only a person of the workspace answers it in the web app. The door is on every plan; during early access people request access.

## Connect it

Add Agent host to your Geenius OS workspace and connect your AI to Geenius OS once, by the steps in the Geenius OS docs. Your AI then has a tool called `geenius_agenthost`.

## How a call looks

The tool is `agents`, and the action is one of its verbs.

```json title="List the agents"
{ "tool": "agents", "action": "list", "args": {} }
```

## Your first ask

```text title="A first ask"
Use Geenius OS and Agent host. List my agents, tell me which are running, and read the last digest of the research agent. Do not send anything to an agent yet.
```

## The tool

::mcp-tools

Ask for the actions with `action: "help"`, and for one action with `of`.

## Habits that work

- **Look before you send.** `get` shows the state and the newest definition.
- **Expect a pending approval.** `send` may answer with an id instead of a result.
- **Read digests, not logs.** `digest` is the last scheduled result.
- **Watch usage.** `usage` shows credits and tool calls against the plan.

## What your AI cannot do

It cannot approve or deny a request, create or change an agent, or read a connector secret. Those are for people, in the app.