# Running an agent for a client

Point an agent at a client's workspace with "Works for", connect that client's own apps to it, and the agent works from their data instead of yours. Here is how the separation holds, and what happens when an app is missing.

If you run work for other people, you usually want one agent per client, reading that client's own Slack, calendar or issue tracker. GPTree does that with workspaces: a client gets a workspace, that client's apps are connected to it, and an agent is pointed at it.

## Point the agent at a client

Open the agent and set **Works for** to the client's workspace. Leaving it on your own workspace keeps the agent personal, which is what a single-operator setup wants. Picking between several client workspaces is a Team plan feature; on other plans the field stays on your own workspace.

Changing "Works for" moves the agent's ongoing thread to the new client, so its briefs, the knowledge it reads and the approvals it needs all follow that client from then on.

## Connect the client's own apps

Connect each app to the client's workspace rather than to your personal account. A connection made on the workspace belongs to the client, and it is the one the agent will use.

This is the part that keeps clients separate. An agent that works for a client uses only that client's connections. It does not fall back to your personal Slack or calendar if the client's is missing, because falling back would mean quietly reading one client's data while doing another client's work.

## What happens when an app is not connected

The agent says so. Instead of guessing, or filling the gap from your own account, the report states that it could not read that source, and you get a prompt to connect it for that client. A brief that could not reach the calendar will say that plainly rather than describing a day it never looked at.

## Blocking a tool for one client

A workspace can also say which tools are allowed inside it. If you block a tool on a client's workspace, an agent working for that client cannot use it, not even on an unattended run, and not even to read. The agent's report says it was blocked rather than going quiet, so you can tell a policy refusal apart from a missing connection. Workspace owners and admins set this under Connect tools, in Approval rules: pick the client's workspace, then add a policy for that tool in deny mode.

## What still belongs to you

The agent itself, its routines and its reports stay with you, the operator. Clients do not get a login to watch their agent. If you need to share output, send them the report the agent produced.

Cost also stays on your account. Agent work spends your credits, and you can separate it per client by giving each client their own project, which is what the usage records are grouped by.
