Build Your AI Phone Agent by Chatting with Claude
You can set up a Daber AI phone agent without opening a single settings screen. Connect the AI assistant you already use — Claude, ChatGPT, Claude Code or Cursor — to your Daber AI workspace with one link, then describe your business in plain language. The assistant writes the agent's instructions, builds the call flow, adds your facts, places a test call and fixes what you didn't like. Most businesses have a working agent in about 15 minutes.
This works through MCP (Model Context Protocol), an open standard that lets an AI assistant use another product's tools. Daber AI runs an MCP server, so your assistant can do in a chat what you would otherwise click through in the dashboard.
What you need
- A Daber AI workspace, and to be its owner or admin. The connection is approved by signing in to your account, so the workspace has to exist first — reserve your launch or talk to us if you don't have one yet.
- An AI assistant that supports custom connectors. claude.ai offers them on every plan, including Free, which is limited to one custom connector (Claude Help Center). ChatGPT supports them under Settings → Connectors, though they may need to be enabled on your plan. Claude Code and Cursor work on any plan.
- The address, which is the same for everyone:
https://api.phone.daber.ai/v1/mcp
Nothing to install and no key to copy. Your dashboard also shows this address, with a copy button, under Settings → Integrations → Connect an AI assistant (MCP).
Step 1 — Connect (2 minutes)
In claude.ai or Claude Desktop:
- Open Customize → Connectors, click +, then Add custom connector. (On a Team or Enterprise plan, an owner adds the connector under Organization settings → Connectors first.)
- Paste the address. Leave every other field empty — no client ID, no secret.
- Click Connect. A Daber AI page opens: click Sign in with Daber AI, log in the way you normally do, pick your business and click Approve.
- Back in a chat, turn the connector on from the tools menu.
In ChatGPT: the same address, under Settings → Connectors.
In Claude Code, one command, then /mcp inside Claude Code to sign in:
claude mcp add --transport http daberai https://api.phone.daber.ai/v1/mcp
In Cursor, add the address to .cursor/mcp.json with an API key from Settings → Integrations → API keys in your dashboard:
{
"mcpServers": {
"daberai": {
"url": "https://api.phone.daber.ai/v1/mcp",
"headers": { "Authorization": "Bearer dbr_live_..." }
}
}
}
To check it worked, ask: "What Daber AI workspace am I connected to?" You should get your business name back.
Step 2 — Build the agent (5 minutes)
If you have a call script, paste it. This gives the best result, because your wording is kept:
Here is our call script. Build me an agent that follows it. Keep the greeting and the closing word for word.
The assistant turns the document into a step-by-step call flow, creates the agent, links the two and picks voice models that can follow it.
If you don't have a script, describe the business instead:
Create a phone agent for my business.
- Business: a dental clinic with two dentists
- It answers: incoming calls when we're busy or closed
- It should: take appointment requests, give our hours and prices, take a message
- It must never: give medical advice or promise a specific dentist
- Tone: warm, brief, professional
- Language: English, and switch to Spanish if the caller does
- If the caller is upset or asks for a person, transfer to: +1 …
Draft the prompt, create the agent and show me the greeting it will say.
Then ask "Show me exactly what it will say when it picks up" and read it out loud. If it doesn't sound like you, say so — that's step 5.
Step 3 — Give it your facts (2 minutes)
An agent should never make up your prices or hours. Hand it the real ones:
Add these to the knowledge base: we're open Monday to Friday 9 to 6. A first consultation costs $80 and takes 30 minutes. Parking is in the building. We don't do emergency care — send those callers to the nearest ER.
What the assistant adds arrives as a draft. Approve it under Knowledge in your dashboard; until you do, the agent won't use it. That is deliberate: nothing reaches your callers without a person saying yes.
Step 4 — Test it (3 minutes)
Ask the assistant to test the agent. It can run the call flow through every path to check nothing dead-ends, and chat with the agent over text to check its answers. To hear it on your own phone, ask:
Call me at +1 … and pretend I'm a new customer asking about prices.
Calling your phone needs an outgoing phone line on your workspace. Without one, test the agent from your browser in the dashboard instead. Afterwards, ask "How did that call go? Show me the transcript."
Step 5 — Fix it by complaining (2 minutes)
Say what was wrong, in plain words:
It talked too much — two sentences maximum. It said "certainly" every time; that's not how we talk. It never asked for a callback number; always ask before hanging up.
The assistant rewrites the agent's instructions. Test again. Three or four rounds is normal.
Nothing here is risky: every change to an agent or a flow is saved as a new version, and the dashboard's Version history rolls back in one click.
Step 6 — Go live
- Ask the assistant: "Set the agent to active."
- Attach a phone number in the dashboard, under Settings → Telephony and the agent's Channel tab.
Connecting a real phone line is deliberately something the assistant cannot do on its own. It is the one step that puts the agent in front of your customers, so a person does it.
Lost? Ask "what's my next step?"
At any point — tomorrow, or in a brand-new chat with no memory of the last one — ask "What's my next step?" The assistant checks your actual workspace and tells you where you are: no agent yet, flow not linked, knowledge base empty, not tested, still a draft, or ready for a phone number. It also flags anything waiting on you, such as unapproved knowledge-base drafts.
Prompts worth keeping
| When you want to… | Say this |
|---|---|
| Check the connection | "What workspace am I connected to?" |
| See what you have | "List my agents and tell me which ones are active." |
| Tighten the writing | "Rewrite the agent's prompt to be shorter and warmer — it's for phone calls, not email." |
| Fill in the gaps | "Fill in any empty persona and behavior fields for this agent." |
| Add facts | "Add our refund policy to the knowledge base: …" |
| Review real calls | "Show me the last five calls and what went wrong in them." |
What the assistant can and can't do
| It can | It can't |
|---|---|
| Create and edit agents | Attach a phone number |
| Turn your script into a call flow | Approve its own knowledge-base entries |
| Rewrite prompts, greetings and objection handling | Change billing or add users |
| Add knowledge-base drafts | See another business's data |
| Test the agent and read transcripts | Do anything your connection wasn't approved for |
Is it safe to let an AI assistant edit my phone agent?
Yes — not because the assistant is careful, but because the system doesn't depend on it being careful:
- Facts need a person. Knowledge-base entries stay drafts until you approve them.
- Everything is undoable. Every agent and flow change is versioned, with one-click rollback.
- It only sees your workspace. The connection is bound to the one business you approved.
- You can cut it off instantly. The connection appears in Settings → Integrations as its own key, named after the assistant. Revoke it and every connected assistant loses access at once. An unusual burst of calls suspends it automatically.
Frequently asked questions
Do I need to know what MCP is? No. You paste one link into your assistant's settings and sign in. After that you just talk to it.
Which AI assistants work? Claude (web and desktop), ChatGPT, Claude Code and Cursor, and any other assistant that can connect to a remote MCP server.
Which plans include it? Every plan, from the $3 launch up — see plans. Test calls to a real phone use call minutes, like any other call.
Can the assistant break my live agent? Every change is saved as a new version, so you can roll back in one click from the dashboard, and facts only reach callers after you approve them.
Prefer to click? Everything here can also be done in the dashboard. Questions — contact us. Also in Hebrew.