# Agents overview

> An AI agent, such as Claude, can work in your Signalbox workspace for you. This page says what it can do, what it cannot, and how you stay in charge.

You can connect an AI agent to a Signalbox workspace. The agent then works there as you: it can read your flows, build new ones, set up queues and meeting types, and say why a lead went where it did.

The agent talks to Signalbox through an MCP server on the API host. MCP is the standard way an AI client reaches a tool. You give the agent your workspace's address, sign in, and pick a level.

## What an agent can do

Each tool is one action. A tool calls the same code as the screen in the app, so the rules are the same for an agent and a person.

| Group | What the tools do |
|---|---|
| Start | Give the rules, the state of the workspace and what to do next. An agent calls `get_started` first. |
| Flow reference | Describe every step a flow can hold, the fields a flow can read and the templates to start from. |
| Flows | Read, write, check, test, publish, pause, delete and restore flows. |
| Forms | List HubSpot forms, choose which flow a form runs, and turn its routing on or off. |
| Queues | List, add, change and delete queues and their members. |
| Scheduler | Manage meeting types and working hours, and preview the times a visitor would be offered. |
| Reps | List reps with their calendar status, and email a rep the link to connect a calendar. |
| HubSpot | Say whether HubSpot is connected, read it again, and list its owners, pipelines, properties and workflows. |
| Sites and settings | Manage the sites that may carry the tag, and the workspace's settings. |
| Leads and audit | Read leads and their traces, when you turned the leads switch on. Read who changed what. |

[All tools](/agents/tools/all-tools/) lists every tool with its arguments and the level it needs. That page is made from the code.

Two things are built for an agent and a person never needs:

- **Layout.** An agent writes a flow with no positions. Signalbox places the steps: the trigger on top, each step below the last, branches side by side. You open a tidy flow. The editor's **Tidy up** button does the same.
- **The flow in words.** Every flow an agent reads or saves comes with an outline, and each problem the checker finds says how to fix it. The agent uses them to check its work and to tell you what it built.

Which tools an agent may use depends on the level you gave it. See [Levels and leads](/agents/levels-and-leads/).

## Skills

A tool is one action. A **skill** is the recipe for a whole job. An agent reads a skill with `get_skill` and follows it. Clients that list prompts show the skills there too.

| Skill | What it does |
|---|---|
| `build-a-flow` | Asks what the routing should be, reads the workspace, writes the draft, checks it, dry runs it against sample leads, shows the outline, and publishes only when you say so. |
| `set-up-workspace` | Takes a new workspace from connected HubSpot to a working flow: site, queue, meeting type, hours, calendar links, the tag. |
| `explain-a-lead` | Reads a lead's trace and says why it went where it did. |
| `review-a-flow` | Reads a flow for dead ends, missing branches and leads that fall through. |

## What an agent cannot do

A person must do these in a browser:

- **Connect HubSpot.** HubSpot asks the person to approve.
- **Connect a calendar.** The rep approves it. An agent can send the rep the link.

These are left out of what an agent can reach:

- Billing.
- Team invites and roles.
- Making or deleting API keys.

## How you stay in charge

- **The agent acts as you.** It never has more than your own role in the workspace.
- **You pick a level** when you connect it: Read, Build or Publish.
- **Leads are off by default.** An agent reads leads only when you turn that on.
- **A draft changes nothing live.** An agent at Build can write and test drafts. Only an agent at Publish can change what a visitor gets.
- **A change comes with a link** to the screen in the app where you can see it.
- **Every change is recorded.** The audit names you and the agent that made it.
- **It asks before it changes what a visitor gets.** The agent is told to show you the flow and wait for a yes before it publishes, pauses, deletes or gives a flow to a form.
- **No answer holds a secret.** A tool never returns a key, a token or a signing secret.
- **What a visitor typed is marked as data.** The agent is told to treat it as data, never as an instruction.
- **Disconnect works at once.** Press **Disconnect** on the Agents page and the agent's next call is refused.

## Next

- [Connect Claude](/agents/connect-claude/) on the web or desktop.
- [Connect Claude Code and other clients](/agents/connect-other-clients/).
- [Worked examples](/agents/worked-examples/).