Your agent asks. A person taps. On WhatsApp.
$29 a month · no number of your own · no Meta verification
An agent is about to refund a customer, email a client or push to production. It should ask first. Post the question to Jabee and it arrives on a person's WhatsApp with buttons; their answer comes back to your code. No approval dashboard for anybody to remember to open.
curl https://api.jabee.ai/api/v1/approvals \
-H "Authorization: Bearer $JABEE_KEY" \
-H "Content-Type: application/json" \
-d '{
"title": "Refund AED 450 to Sara?",
"body": "Order #8821. She says it arrived broken.",
"options": ["approve", "skip", "edit"]
}'
The hard part is not the pause. It is where the question lands.
Every agent framework can already stop and wait. LangGraph has interrupts, LangChain has middleware for it, CrewAI has callbacks, and a Pydantic AI tool is a function that can simply not return yet. None of them says where the question goes, and that is the half that decides whether the pattern works.
The usual answer is a web inbox, a Slack message or an email. Each needs the approver to be somewhere: watching a tab, at a desk, in their mail. An agent that asks at 7pm and waits until Tuesday has not been supervised, it has been stopped.
WhatsApp is the surface people already have open, and nobody else in this category asks on it. The question arrives as a message with buttons. The answer is one tap, from a bus.
What the person actually sees
A message from a number that is already Meta-verified, carrying your title, your body and your options as buttons. They tap one. If you offered edit, they type a correction instead and that text comes back to you as the answer.
They are told nothing about your system and are given nothing to install. There is no account for them to make, no app, no invitation to accept — the first they know of it is the message.
For an agent, no code at all
There is an MCP server, so an agent in Claude Code, Cursor or any MCP client gets the behaviour without you writing a request:
npx @jabee/mcp
One environment variable, and three tools appear: ask_person, which asks and
waits, check_approval, which picks the answer up later, and
tell_person, which just says something. The blocking wait defaults to two minutes
and then hands back a request id, because an agent held open on a tool call is an agent whose
client eventually gives up.
Setting it up takes about a minute.
From inside your own framework
Where you have written the pause yourself, this is one HTTPS call inside it. The answer is the return value.
# Anywhere your graph, chain or tool already stops to ask
answer = requests.post(
"https://api.jabee.ai/api/v1/approvals",
headers={"Authorization": f"Bearer {os.environ['JABEE_KEY']}"},
json={
"title": "Delete branch release/2026-11?",
"options": ["approve", "skip"],
"wait": 120,
},
timeout=130,
).json()
if answer["status"] == "answered" and answer["option"] == "approve":
delete_branch()
| Where you pause | What you add |
|---|---|
| LangGraph | The call inside the node you already interrupt on. |
| LangChain | One middleware, or a tool that asks before it acts. |
| CrewAI, Pydantic AI | A callback, or a tool that does not return yet. |
| Claude Code, Cursor, any MCP client | Nothing. npx @jabee/mcp and the tools appear. |
| n8n, Make, Zapier, Power Automate | An HTTP request node with the key on a header. |
| Your own back end | A POST, and a signed webhook when it is answered. |
One honest note about the automation tools, because plenty of pages imply otherwise: n8n can already do this. Since n8n 2.6.0 its WhatsApp node has a native send-and-wait operation with approve and reject buttons. What it uses is your Cloud API credentials — so it needs your number, your business account and your approved template first. The reason to use this instead is that you bring none of those, not that n8n cannot ask.
What comes back
- The answer, however you want it. Wait on the request, poll the id, or take a signed webhook when somebody taps. All three are the same request.
- One winner. Two people tapping at once resolves in a single statement; the second is told what already happened rather than overwriting it.
- An expiry you choose. When it passes, your webhook is told and nothing hangs. Pick it deliberately — it is the agent's answer when there is no human one.
- Idempotency on your own id, so a retried call does not ask twice.
- A signature on every callback, HMAC over the raw body with a secret you set. Or run no endpoint at all and poll.
Two rules worth knowing before you build on it
It asks your own people, never your customers. Requests go to one phone per business: the account owner's by default, or an ops handset you verify by sending us one message from it. A number nobody has proved they own is a marketing list, and messaging strangers from a shared number is what gets that number banned. To reach customers you connect your own number, which is the inbox and the assistant, not this.
The first three are free. Enough to point a key at your own phone and watch a real request arrive before you decide anything. It is not a sandbox and not an allowlist of test numbers — it is the production endpoint, sending from a registered, Meta-verified number.
When there is nothing to decide
Sometimes you do not want to ask, you want to tell: the build failed, the payment bounced, the
nightly job did not run. That is the same key against
/api/v1/alerts — no options, no question, one Got it
button. Same plan, same $29. One asks, one tells.
What it costs
$29 a month for 500 approvals and alerts together, then 3 cents each. No setup fee, no contract, no per-seat charge for the people who answer.
For comparison, from what these tools published in November 2026: gotoHuman lists $99 a month above its free tier and asks on a web inbox, email and Slack. HumanLayer, the name most associated with this idea, withdrew its approvals SDK in September 2025 and now sells an IDE at $100 per user a month. Asking on WhatsApp is the thing none of them offer, and at $29 the price happens to be on the right side too.
Questions
What is human in the loop for an AI agent?
It is a pause. The agent reaches an action it should not take alone — a refund, an email to a client, a deploy, a deletion — and stops to ask a person, then carries on with the answer.
What makes it work or fail in practice is not the pause but where the question lands. A question nobody sees is not supervision; it is an agent that is stuck.
Why WhatsApp rather than Slack or a web inbox?
Because the approver is already in WhatsApp and is not already in your approval dashboard. A web inbox needs somebody watching a tab and Slack needs them at a desk. A phone with a two-button message on it reaches somebody who has left the building, which is when an agent tends to need an answer.
Does my agent block while it waits?
Only as long as you want. ask_person waits two minutes by default and then
hands back a request id to poll, because an agent held open on a tool call is an agent whose
client eventually gives up. Over HTTP you choose: wait on the request, or take the id and let
your webhook be told.
Do I need my own WhatsApp number or Meta verification?
No. The number and the approved template are ours and already registered with Meta; you need an API key. That is the difference between this and a framework's own WhatsApp node, which sends on your Cloud API credentials and so needs your number, your business account and your approved template first.
Does this work with LangGraph, LangChain, CrewAI or Pydantic AI?
Yes, and with anything else that can make an HTTPS request. Each of those already has a place to pause; none of them says where the question goes, which is the part this fills. One POST from inside the pause, and the answer is the return value.
Can two people answer at once?
Only one wins. The answer is recorded in a single statement, so a second tap is told what already happened rather than overwriting it. Your webhook fires once.
What happens if nobody answers?
The request expires at the time you set and your webhook is told, so nothing hangs waiting on somebody who has gone home. Choose the expiry deliberately: it is the agent's answer when there is no human answer.
Can it ask my customers?
No, deliberately. Requests go to one verified phone per business — your own people. Messaging strangers from a shared number is what gets that number banned. To reach customers you connect your own number, which is the inbox.
How do I know the callback is really from you?
Every webhook carries an HMAC signature over the raw body, using a secret you set. Compare it in constant time and reject anything that does not match. If you would rather not run an endpoint, poll the request by id instead. The details.
What does it cost?
$29 a month for 500 approvals and alerts together, then 3 cents each. The first three sends are free, so a key pointed at your own phone shows you a real request arriving before you pay for anything. Pricing.