# Build a business from a chat

> A worked MCP session: from "set up a front office for Bright Smile Dental" to a live receptionist, trained knowledge, an SMS text-back workflow, and A2P registration — no dashboard.

This is the transcript-level shape of a real session. You say: *"Set up a front office for Bright Smile Dental — dental practice, brightsmiledental.com. Phone, chat, texting. Missed calls should get a text back."* The AI does:

```text
1.  create_project        { name: "Bright Smile Dental" }
2.  create_service        { type: "chatbox",      name: "Website chat" }
3.  create_service        { type: "receptionist", name: "Front desk" }
4.  create_service        { type: "sms",          name: "Texting" }
5.  create_service        { type: "crm",          name: "Patients" }
6.  add_knowledge         { kind: "url",  url: "https://brightsmiledental.com" }
7.  add_knowledge         { kind: "qa",   question: "Do you take Delta Dental?",
                            answer: "Yes — Delta, Cigna, and Aetna PPO." }
8.  setup_receptionist    { greeting: "Thanks for calling Bright Smile Dental!" }
9.  search_phone_numbers  { area_code: "512" }
10. buy_phone_number      { number: "+1512…", sms: true, label: "Main line" }
11. save_sms_compliance   { business_name, ein, address, opt_in_method, … }
12. register_sms_compliance
13. create_workflow       { name: "Missed call → text back",
                            graph: notify-SMS node with the wording }
14. set_workflow_triggers { events: ["call_missed", "call_voicemail"] }
15. test_workflow         { phone: "+1512…yours" }   → the text arrives
16. publish_workflow
17. get_analytics         → confirm everything reports
```

## What to check as you go

- **After step 6** — `list_knowledge` shows the crawl's pages; add Q&A pairs for anything the site is vague about.
- **After step 12** — `get_sms_compliance` reports status; registration approval can take days, so everything else proceeds while it's pending.
- **Step 15 before 16** — test with your own phone before publishing; the run log shows exactly what fired.

## Patterns that work well over MCP

- **The agency clone** — "set up the same stack as Client A for Client B, with their name and website" — the AI reads one business with `list_*` tools and re-creates it.
- **The weekly readout** — a scheduled assistant calls `get_analytics` across businesses and writes the Monday summary.
- **Bulk contact ops** — `list_contacts` + `add_contact` to sync a list, respecting DND flags the tools return.

> **GIVE THE AI THE PLAYBOOK:** Paste the relevant [vertical playbook](/docs/playbooks/overview) into the conversation before you start. The model then builds the flows that vertical actually needs instead of generic ones.