Quickstart
Create an agent, give it a tool, and watch it reply — all in the sandbox, before a single WhatsApp message is sent. About five minutes.
Before you start
- An account and an API key from the dashboard — use a
mcp_sk_test_…key here so nothing sends or bills. See Authentication. - An HTTPS endpoint the agent can call as a tool (optional for this walk-through).
1. Create an agent
An agent needs a name, a goal, and a persona. The goal is plain language — it’s what the agent is there to do.
2. Give it a tool
A tool is one of your endpoints, described so the model knows when to call it. Keep the description action-oriented; the full schema and signing details are in Building tools.
3. Test it in the sandbox
The sandbox runs the entire pipeline — retrieval, tool calls, the model — but sends no WhatsApp message and persists nothing. Tool calls are made with a test flag so your endpoint can branch if it wants to.
4. Connect a number and go live
Connecting a WhatsApp number is done with our team through the dashboard — a Meta pre-flight checks the number and token before anything goes live, so a bad credential fails at setup, not at message time. Once the number is bound to your agent, real customer messages flow through the same pipeline you just tested.
Next steps
- Building tools — the signed envelope, argument mapping, and importing tools from your existing docs.
- Customer identity — let tools know which of your customers is asking.