@jabee/mcp
Sends WhatsApp notifications and approval requests to verified team members through Jabee's approved sender. Supports one-way completion messages via tell_person, decision requests via ask_person with buttons, optional free-text amendments, expiry, and idempotent external IDs, plus check_approval for reading back pending or cancelled requests. It cannot read personal chats or message customers.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@@jabee/mcp ask Sam on WhatsApp before you force push to main"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
@jabee/mcp
Your agent messages you on WhatsApp when it is done — and asks you first when it is about to do something you would want a say in.
JABEE_API_KEY=jak_... npx @jabee/mcpTwo jobs, one key, no phone number of your own.
Three different things are called a WhatsApp MCP server
One of them can cost you your number, so it goes first.
Kind | What it connects to | Can read your chats | Needs your own number | The catch |
Unofficial client — QR-linked, usually built on whatsmeow or Baileys | Your personal WhatsApp, as a linked device | Yes | Yes — your own, personal one | WhatsApp's terms do not permit it, and the number can be banned |
Your own Cloud API | A WhatsApp Business account you registered | No — only conversations on that business number | Yes, plus business verification and approved templates | Days of setup before the first message, and a template review each time the wording changes |
A borrowed sender — what this is | Jabee's approved sender | No | No | It asks and tells your own verified people. It cannot reach your customers |
Most of the repositories you find by searching whatsapp mcp are the first
kind: a Go or Node client that scans a QR code, registers as a linked
device, stores your history in a local SQLite file and exposes it as MCP
tools. They are genuinely impressive, and several of them say the risk out
loud in their own READMEs. The risk is not the software, it is the account —
WhatsApp's terms do not permit third-party clients, and enforcement lands on
the phone number, which is the number your family messages you on.
That is also the trade: a personal account is the only way to read your own chats from an agent. No official route can do that, including this one. If that is what you came for, this is the wrong package.
This is not a disinterested opinion and it would be silly to pretend otherwise: Jabee sells the third kind. The rule it keeps for itself is in the open — no unofficial library has ever been in this product. The longer comparison is at jabee.ai/docs/whatsapp-mcp/.
Related MCP server: WhatsApp Channel
Tell me when it is finished
You gave an agent twenty minutes of work and walked away. When it
finishes, tell_person sends you a line on WhatsApp — in the agent's own
words, so it says what it actually did rather than that something
happened.
The usual answer to this is a push service and another app on your phone to install and then ignore. This lands on the one that is already open, and the same key can also ask you a question, which a notifier cannot.
Put this in your CLAUDE.md, or whatever your agent reads, because it
will not call a tool it was not told about:
When a long task finishes, call tell_person with one line on what
changed. Before anything irreversible — a force push, a production
deploy, a migration, deleting a branch, spending money — call
ask_person and wait for the answer.Ask me before you do that
The other half, and the reason this is not just a notifier. Your agent is
about to issue a refund, send a quote, post something publicly, or email
a list. ask_person asks somebody first, waits, and hands back the
answer. The person gets buttons on their phone.
Why this rather than building it
You can build this on Twilio in a weekend, and several people have. What is in the weekend: a number, a message template Meta has to approve, a webhook, a state table, and a decision about what happens when two people tap at once or nobody taps at all.
This needs no phone number of your own, because the number and the template are Jabee's and already approved. It also already handles the parts that are easy to get wrong:
Two people answering at once is settled in one statement, so exactly one wins and the second is told what happened.
A request nobody answers expires rather than hanging your agent for ever.
An Edit button returns free text, because the answer is often "yes, but change this" rather than yes or no.
Sending the same
external_idtwice returns the request you already made, so a retry does not ask somebody the same question twice.
Setting it up
Make a key under Management, then Approvals, at
app.jabee.ai. It starts jak_.
Then, in your MCP client's configuration:
{
"mcpServers": {
"jabee": {
"command": "npx",
"args": ["-y", "@jabee/mcp"],
"env": { "JABEE_API_KEY": "jak_..." }
}
}
}The person who answers needs a verified phone, which they get by sending one message from a link in the dashboard. Until they do, requests fall back to email.
The tools
tell_person
One-way, and the one to reach for first. A build failed, an order
shipped, a run finished. Takes the same title and body as
ask_person, reaches your own verified team and never your customers,
and returns as soon as it is sent rather than waiting for anybody.
ask_person
Asks for a decision and waits for it.
| The decision, in one line. Required. |
| Why you are asking. |
| A link to the thing being decided. |
| Offer an Edit button that takes free text. |
| Your own id, so a retry is safe. |
| How long the decision stays useful. |
| How long to block. Defaults to 120. |
Returns the decision, or the request id if nobody has answered within
wait_seconds. Two minutes is the default because an agent blocked on a
tool call is an agent whose client may time out.
check_approval
Reads a request back by id, for when ask_person returned while still
pending. It also reports a request that was cancelled ("Do not act on it")
and any outcome recorded against it, so an agent that comes back later does
not act on a decision somebody has since withdrawn.
The Claude Code channel
The same package ships a second entry point: a channel for Claude Code, which puts the permission prompt on your phone.
JABEE_API_KEY=jak_... claude --dangerously-load-development-channels server:jabeeWhen Claude asks to run Bash, Write or Edit, the prompt arrives on
WhatsApp with Approve, Skip and Edit. The terminal dialog stays open the
whole time and the first answer wins — the keyboard or the phone,
whichever you reach first. Claude Code's own channels are Telegram,
Discord and iMessage; this is the one for the app that is already open.
Add it to .mcp.json (or ~/.claude.json with an absolute path) first:
{
"mcpServers": {
"jabee": {
"command": "npx",
"args": ["-y", "--package=@jabee/mcp", "jabee-channel"],
"env": { "JABEE_API_KEY": "jak_..." }
}
}
}Three things worth knowing before you rely on it:
The development flag is required, and that is not our choice. Channels are a research preview, and during it Claude Code registers only channels on an Anthropic-maintained allowlist, or on an organisation's own
allowedChannelPlugins. Without one of those the channel is silently not registered, so--dangerously-load-development-channelsis the honest instruction rather than a footnote. A Team or Enterprise admin can allowlist it instead.Tapping Edit denies the command and sends Claude your words, which is usually what you meant: "no, the staging one". Approve allows, Skip denies.
Nobody answering sends nothing. If the request expires or cannot be delivered, no verdict is sent and the prompt is still waiting at your terminal. An invented denial would kill a job no person refused.
It also gives Claude one tool, reply, for answering you on WhatsApp.
Every reply is a real message billed to the account, so it is told not to
narrate progress with it.
One environment variable beyond the key:
JABEE_PERMISSION_EXPIRY_MINUTES (default 15, floor 5). There is nobody
to address — a request goes to the one phone this business has verified,
or the account owner's.
Limits
60 API calls a minute per key, and 200 new requests per business in any
24 hours. A repeated external_id does not count against the daily
figure.
The daily limit is there because every request is a message from a number every Jabee business shares, and one number's standing with Meta is everybody's. If you genuinely need more, email hello@jabee.ai.
What it costs
$29 a month for 500 notifications and approvals together, then 3 cents each. They share a plan because a notification is an approval without a question. The first three are free, so you can watch one land on your phone before paying for anything.
For one notification to one phone, cheaper and free things exist and we would rather say so: CallMeBot's free API and ntfy both do that, and $29 is not worth paying for one line a week. What $29 buys is the case they do not serve — a team: a request that reaches the right person, an acknowledgement, an answer that comes back, and an audit trail of who decided what.
Not for messaging customers
This reaches people on your own team who have verified a phone with us. It cannot message your customers, deliberately: a shared number messaging strangers is what gets numbers banned. To reach customers you connect your own WhatsApp number, and then it is your number and your relationship.
Documentation
The HTTP API underneath this, with signature verification in Node and Python if you would rather use a webhook than an agent.
Without MCP
If your agent does not speak MCP — or you are calling from cron, CI or a
web handler rather than from an agent — the same four jobs are a plain HTTP
call. examples/ has them in Node and in Python, with no SDK
and no dependencies: send an alert, ask and wait for an answer, check in a
scheduled job so silence raises the alarm, and receive the answer by signed
webhook. Both sets have tests that run against a stub, so you can check them
without a key.
About this repository
This repository is the published source of @jabee/mcp 0.1.3 — the
package npx installs, copied here so you can read it before running it
against your own account. It is not the Jabee service: the WhatsApp sender
and the Meta-approved template are the parts that cannot be installed.
Issues and questions belong here, or hello@jabee.ai. If the version above is behind the one on npm, the npm one is current and this copy is stale — say so in an issue and it will be updated.
Available Tools
3 toolsask_personAsk a person to approve something on WhatsAppA
Ask a named person to approve, reject or amend something, on WhatsApp, and wait for their answer. Use this before doing anything that is expensive or impossible to undo: issuing a refund, sending a quote, posting publicly, emailing a list, deleting data. The person gets buttons on their phone and can also reply with free text to amend what you proposed. Returns their decision, or the request id if nobody has answered yet.
Inputs: title (required), url, body, external_id, +3 more.
check_approvalCheck whether a request has been answeredA
Read the current state of a request you asked earlier, by its id. Use this when ask_person returned while still pending.
Inputs: id (required).
tell_personTell a person something on WhatsAppA
Send a one-way notification to your own phone on WhatsApp: a build failed, an order shipped, a job finished. This reaches your own verified people, never your customers. Returns once it has been sent; it does not wait for a reply.
Inputs: title (required), url, body, external_id.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 3 tools
The three tools have clearly distinct roles: ask_person initiates a blocking human approval, check_approval polls an existing request, and tell_person sends a one-way notification. The descriptions make the waiting/blocking versus notification-only distinction explicit.
All tool names follow a consistent snake_case verb_noun pattern: check_approval, ask_person, tell_person. There are no mixed conventions or vague standalone verbs.
Three tools are well-scoped for a focused human-in-the-loop WhatsApp approval/notification server. Each tool earns its place: initiate approval, check pending approval, and send non-blocking notification.
The core approval lifecycle is covered: create a request, poll its state, and notify without waiting. Minor gaps exist around listing all pending approvals or canceling/revoking a request, but these are not fatal for the stated workflow.
Maintenance
Related MCP Connectors
Your own WhatsApp in Claude and ChatGPT: read chats, draft replies, send messages you approve.
WhatsApp for Claude, ChatGPT or any AI agent: search, reply and send. Personal or Business API.
WhatsMCP connects Claude and other MCP-compatible AI agents directly to WhatsApp. Send and receive text, images, documents, and voice notes; manage groups (create, add/remove members, promote admins); look up contacts and profiles; follow channels; and read call and message history — all through a standard MCP interface. For voice use cases, WhatsMCP offers SIP-based calling plans (inbound-only, or full inbound/outbound) so AI voice agents can answer and place WhatsApp calls, plus low-latency WebSocket integrations with voice agent providers like ElevenLabs. Multiple WhatsApp accounts can be paired and managed per workspace, with webhook support for real-time inbound message delivery to your own infrastructure.
WhatsApp for your AI agent, sent from your own number. Source-available (Elastic License 2.0).
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceHuman-in-the-loop approvals and notifications for AI agents via WhatsApp. Enables Cursor, Claude Code, and autonomous AI agents to reach users away from their computers.112 npmISC
- FlicenseNot gradedqualityAmaintenanceEnables a Claude Code session to send and receive WhatsApp messages from a linked personal account or the official Cloud API, including photos, voice notes and documents, with local voice transcription and allowlist-based access control. It also lets users approve or deny Claude's tool requests remotely by emoji reaction and keeps per-chat personalities, cron jobs and conversation context across restarts.-
- FlicenseNot gradedqualityAmaintenanceEnables users to drive a Claude Code session from WhatsApp, with bidirectional messaging, media and voice-note handling, access-controlled chats, per-group personalities, and remote approval of tool requests via emoji reactions.-
- FlicenseNot gradedqualityAmaintenanceEnables a Claude Code session to send and receive WhatsApp messages through either a linked-device pairing or the official Cloud API, including media, local voice-note transcription, and @-mentions. It also gates inbound chats with pairing codes, allowlists and per-group policies, and lets users approve or deny tool requests with an emoji reaction.-