# Connect Claude or ChatGPT to your workspace

> Add the Broker & Agent MCP server to Claude, ChatGPT or any MCP client, sign in, and ask your CRM questions in plain English.

Broker & Agent runs an MCP server (Model Context Protocol, the open standard AI assistants use to connect to other software). Add it to your assistant once, sign in, and your workspace answers questions directly:

- "What does my day look like?"
- "Who went quiet this month?"
- "What did Dana email me yesterday?"
- "Add a note to Dana Reyes: wants a finished basement."

The server address is the same for every client:

```text
https://mcp.brokerandagent.com/mcp
```

It works on any Workspace plan, for anyone in the workspace from agent up; transaction coordinators can't connect it. Each tool sees exactly what you see in the app, nothing more.

## Claude

Works in Claude on the web and in the Claude desktop app.

1. In Claude, open **Settings → Connectors** and choose **Add custom connector**.
2. Name it **Broker & Agent** and paste the server address above. Choose **Add**.
3. Choose **Connect**. Claude opens Broker & Agent: sign in and approve the connection.
4. In a new chat, ask something like "What's on my calendar today?"

> [!NOTE]
> On Claude Team and Enterprise plans, an owner adds the connector for the organization first. Each person then connects it with their own Broker & Agent sign-in.

## ChatGPT

1. In ChatGPT, open **Settings → Apps & Connectors**. Under **Advanced settings**, turn on **Developer mode**.
2. Choose **Create**, name it **Broker & Agent**, paste the server address, and choose **OAuth** for authentication.
3. ChatGPT opens Broker & Agent: sign in and approve the connection.
4. In a new chat, add the connector from the tools menu and ask your question.

## Claude Code and other MCP clients

Clients that support OAuth find our sign-in automatically. Clients that don't can send an [API key](https://brokerandagent.com/developers/api/authentication) instead.

```bash
claude mcp add --transport http brokeragent https://mcp.brokerandagent.com/mcp
# then run /mcp inside Claude Code to sign in
```

```bash
claude mcp add --transport http brokeragent \
  https://mcp.brokerandagent.com/mcp \
  --header "Authorization: Bearer bak_YOUR_KEY"
```

```json
{
  "mcpServers": {
    "brokeragent": {
      "type": "http",
      "url": "https://mcp.brokerandagent.com/mcp",
      "headers": { "Authorization": "Bearer bak_YOUR_KEY" }
    }
  }
}
```

The server speaks streamable HTTP. A request without a credential gets a `401` that points OAuth clients to our authorization server.

## What the assistant can do

Reading is generous, so it can answer questions about your book well: contacts and their timelines, your pipeline and reports, tasks, appointments and your calendar, transactions, conversations, documents, smart lists and the team.

Writing is deliberately light: notes, tasks, leads, appointments, and enrolling a contact in an automation.

Some things are left out on purpose. The assistant can't write or send a text or email, delete anything, move pipeline stages, or change settings, billing or the team. That work stays in the app, where you can see it happen.

> [!NOTE]
> Two of the writes can lead to messages going out. Sending a lead starts your new-lead automation, and enrolling a contact in an automation starts that automation, either of which sends if it's set to. Enrolling needs its own permission, `automations.run`: the connection has to ask for it, and you leave **Automations** at **Read and write** when you approve. Without it, the assistant can't enroll anyone. And with **Ask me before it sends or deletes anything** on (the default), an enrollment waits on your Today for you to approve.

## Privacy and permissions

- **It sees what you see.** Every tool calls the API as you, so a record you can't open in the app, the assistant can't read either.
- **You approve it, and you can take it back.** Each person connects with their own sign-in. Change its access or disconnect it any time in **Settings → Integrations & API**, under **Connected AI apps**; owners and admins see everyone's connections. Disconnecting also removes any webhooks the app set up.
- **The connection only reaches its own tools.** A token approved for the MCP server works for the calls its tools make and nothing else in the API.
- **Your assistant's provider sees what you ask about.** What the assistant reads goes to that provider, such as Anthropic or OpenAI, under your agreement with them.
- **Outside text is marked.** Words from outside your workspace, like a client's email, reach the assistant as data, never as instructions. The few write tools also limit what a misleading message could get the assistant to do.

## Troubleshooting

**It says it can't find a contact I can see.** Make sure you signed in to the right workspace. If you belong to more than one, ask the assistant to list your workspaces and say which one you mean.

**Sign-in opens, then nothing happens.** Allow pop-ups for the assistant's site, then choose **Connect** again.

**"Too many requests."** Each connection can make 120 requests a minute. Wait a minute, then ask again.

---

Source: https://brokerandagent.com/developers/mcp
