You have read that an AI agent can run WhatsApp for you. Before you wire it into your morning, you want the boring truth: which actions it can take, and where each one dead-ends. This is that list, a reference and not a pitch, and half of it is the limits. Every tool below is real and verified against production, and so is every place it stops.
What are the WhatsApp MCP tools an AI agent can actually call?
Nine. An AI agent connected to a WhatsApp MCP server on Blueticks can call nine tools by name: audiences, campaigns, chats, contacts, engine, groups, scheduled_messages, utils, and webhooks. You describe an outcome in plain language and the agent picks the tool.
The whole set, one line each:
- chats - read, search, and reply inside conversations
- scheduled_messages - send now or schedule for later
- contacts - list contacts, fetch a profile picture
- groups - create and manage WhatsApp groups
- audiences - reusable contact lists for sending
- campaigns - paced bulk delivery you can pause
- webhooks - get an HTTPS callback on events
- engine - check, reload, or log out the session
- utils - validate numbers, preview links, tell the time
That is the complete whatsapp mcp tool list. New to the protocol layer? The WhatsApp and MCP primer covers what MCP is and how a client discovers tools; this page assumes that and gives you the surface. And a connector that only reads and replies is only half the loop; these nine tools are the half that acts.
Is it nine tools or ten? (and why both numbers are in circulation)
Nine. The codebase declares ten canonical tool names, but only nine are ever published to a client, so nine is what an agent actually sees. A name can exist in the source before it is exposed on the connector, which is why a count from the code and a count from a live session differ by one.
The practical version: what matters is not how many names live in a file, it is how many tools show up when your agent asks the server what it can call. That answer is nine, and it is nine for every client. Claude and ChatGPT are served the same nine, not a bigger menu for one. Seeing "ten" in one place and "nine" in another is just the source tree versus the exposed surface, and the exposed surface is the one you use.
What can each tool do, and where does each one stop?
Each tool is a small cluster of actions with a hard edge. The table below is the reference: what the tool does in plain outcomes, and where it stops. Read the right-hand column as carefully as the left. The limits decide whether these whatsapp mcp server tools fit your job.
| Tool | What it can do | Where it stops |
|---|---|---|
| chats | Look up chats, search them, read message history, load older history, pull media, list participants, and mark a thread read. | Reads only what your own linked session has synced. It cannot recover messages the session never received, and marking read acts on your side, not the recipient's. |
| scheduled_messages | Send a message now, schedule one for later, list scheduled items, edit or cancel one, and check its delivery ack. | A scheduled send needs a live session at fire time, not at schedule time. The read ack is best-effort and often absent (see the next section). |
| contacts | List your WhatsApp contacts and fetch a contact's profile picture. | It reads your existing contacts. It will not enrich a contact with a CRM record, and it cannot conjure a stranger who was never in your list. |
| groups | Get a group's details (optionally with its members inline), create a new group, and update an existing one. | No leave-or-delete action on this surface, and every membership change still obeys WhatsApp's own group rules and limits. |
| audiences | Build and reuse contact lists as send targets: create, get, update, delete a list, and append, update, or remove contacts in it. | It is a list container, not a consent ledger. It does not record who agreed to hear from you and will not screen out people who blocked you. |
| campaigns | Schedule an audience for paced bulk delivery, then pause, resume, or cancel the run. | Pacing is not permission. There is no template approval and no guaranteed inbox placement, and slow delivery does not make cold bulk sending safe. |
| webhooks | Subscribe to WhatsApp events over an HTTPS callback: create, list, get, update, and delete subscriptions. | It delivers the events the platform actually emits. An event that never fires (a missing read, below) will not arrive just because you subscribed to it. |
| engine | Report the WhatsApp session's status, reload it, or log it out. | It inspects and restarts an existing session. It cannot create a session from nothing, and it cannot keep a browser tab you closed alive. |
| utils | Validate a phone number, generate a link preview, and return the current date and time. | Validation checks format and reachability signals, not whether the person behind the number wants your message. |
A few carry more than the name suggests. The chats tool alone exposes eight actions, and scheduled_messages carries six, including an ack action to check delivery after the fact. When people ask about whatsapp ai agent capabilities, this table is the honest answer: nine tools, roughly three dozen actions, and a real edge on each one.
Which delivery states can you trust, and which are best-effort?
Trust queued, sending, and delivered. Treat read as best-effort. A message moves through a lifecycle you can rely on up to delivery, then one final step, the read receipt, that frequently does not fire even when the recipient has plainly read it. Build around the dependable states and never block a workflow on read.
Here is the chain in plain terms:
- Queued - the message is accepted and waiting. Dependable.
- Sending - it is on its way out from your session. Dependable.
- Delivered - it reached the recipient's device. Dependable.
- Read - the recipient opened it. Best-effort, and often missing.
The read gap is a known, documented issue on our side, and it is fair to say so flatly. The read event does not always get recorded, so an agent that waits for "read" before its next step can wait forever on a message the person read an hour ago. Design any read-reactive automation to tolerate the signal being absent. And if the recipient turned read receipts off, no event is correct behaviour, not a bug.
One thing to keep straight: this surface drives your own WhatsApp account, so the states above are our own lifecycle, not the WhatsApp Business Platform Cloud API's sent/delivered/read contract. Do not assume Meta's message-ID semantics or per-message guarantees apply here. Different product, different rules.
What can no WhatsApp MCP tool do for you?
None of these nine tools will get you consent, keep your number safe, or turn your personal account into a business API. These hard edges hold no matter how you phrase the prompt. Knowing them up front saves you a banned number and a bad week.
Read these as flatly as they are written:
- It will not get you consent. The tools send what you tell them to send. Consent is the sender's responsibility. An agent will happily draft two hundred first-contact messages; whether those people agreed to hear from you is on you, not the tool.
- It will not guarantee you are not banned. Ever. Automating a personal WhatsApp account carries real risk, because unauthorized automated or bulk messaging violates WhatsApp's Terms of Service and Meta enforces it. Read-and-reply on existing threads is low-risk; cold-blasting strangers is high-risk. Anyone promising a no-ban setup is selling something they cannot deliver.
- It does not turn your number into a Cloud API number. This drives your existing account over WhatsApp Web or a hosted engine, so you inherit none of the Cloud API's machinery: no template categories, no per-message pricing, no business verification, and none of its sanctioned-delivery guarantees.
One more thing, plainly: we are not the first WhatsApp MCP. Open-source servers predate us, they are a real option, and the self-hosted-versus-hosted decision already lives in our hosted vs self-hosted comparison. It owns that argument; this page points you there.
Which AI client can use these tools today?
Claude, as a one-click connector, today. Claude is the only AI client with a live in-product connector for this WhatsApp MCP server right now. ChatGPT can reach the exact same nine tools, but through the manual developer-mode route, not an in-product card. There is no Gemini path. That is the whole client picture.
- Claude - a custom connector using remote MCP is live in-product. The click-path for connecting your WhatsApp number lives in how to connect WhatsApp to Claude. Follow it there; I am not reproducing a step of it here.
- ChatGPT - MCP servers are reached through developer mode, a more manual flow than a connector card. It works, and it gets you the same nine tools. The exact route is in connect WhatsApp to ChatGPT.
The takeaway: the client changes the setup ergonomics, not the capability. So the answer to what can claude do with whatsapp and the answer for ChatGPT are the same nine-tool answer. Pick the client you already live in.
Which repeating jobs are these tools actually for?
The tools earn their keep on work that repeats, not a one-time send. The value shows up when the same job runs every week without you starting it: the Monday follow-up sweep, the standing reminder loop, the triage pass every morning. A single tool call is a party trick. A job that repeats is the reason to connect anything.
Think in loops, not one-shots:
- The Monday triage pass. Every Monday, the agent reads the weekend threads with
chats, tells you who is still waiting on a reply, and drafts the answers for you to approve. Same job, every week. - The standing reminder loop. A recurring nudge that fires on a schedule through
scheduled_messages, so the invoice chase or the appointment reminder goes out without you remembering it. The route for that is send WhatsApp reminder messages. - The weekly recurring send. The same message to the same audience on a cadence, which is recurring WhatsApp messages territory, not a one-off blast.
I am deliberately not turning this into a recipe book. Concrete prompt recipes for these jobs already live in how to automate WhatsApp with AI. The point is narrower: match the tool to the job that recurs, where a tool list becomes leverage. As one operator who runs a standing Monday triage framed it, the tool list is a menu you can only order from while the kitchen is open.
What has to be in place before any tool call reaches a phone?
A live WhatsApp session. Every one of the nine tools is a menu of errors until a real session is connected to your account. That is the single prerequisite, and it is where most first attempts fall over. The tools are ready. The question is whether there is a session for them to act through.
Here is the failure mode nobody warns you about. You connect through a browser tab during the day, everything works, then you shut the laptop at 18:00. Your clients message at 21:00. The agent tries to reply and every tool call errors, because the session that lived in that tab closed when the tab did. Nothing is broken. Nothing is connected either.
WhatsApp itself caps how you connect: a standard account allows up to four linked companion devices beyond your primary phone, and any session you run counts against that. So the real choice is not whether to connect, it is whether your session survives you closing the laptop. A browser tab does not. An always-on hosted engine does, and that is the honest reason to pay for anything here: it keeps the session up while you sleep, so a 21:00 request actually leaves your number. No developer console, no template queue, no per-message fees, just a session that does not sleep.
You now know exactly which nine tools exist and exactly where each one stops. The only open question left is whether there is a live session for your agent to run against, so put your own number behind an always-on engine and point the agent at the job that repeats every week, not the one-off you would have done by hand anyway.
FAQ
What tools does the WhatsApp MCP server have? It has nine: audiences, campaigns, chats, contacts, engine, groups, scheduled_messages, utils, and webhooks. Together they cover reading and replying in chats, sending and scheduling messages, managing contacts, groups, audiences and campaigns, event webhooks, session control, and utilities like phone validation.
What can Claude do with WhatsApp? Claude can read and search your chats, draft and send replies from your own number, schedule messages, manage contacts and groups, run paced campaigns, and subscribe to event webhooks. It does all of it by calling the nine tools on your behalf. It cannot get you consent, and it cannot guarantee your number is never banned.
How many WhatsApp MCP tools are there? Nine are exposed to any client. The codebase declares ten canonical names, but only nine are ever published to the connector, so nine is what an AI agent actually sees when it lists what it can call. Both Claude and ChatGPT are served the same nine tools, with no smaller or larger menu for either.
Can an AI agent read my WhatsApp messages? Yes, through the chats tool, but only the messages your own linked session has synced. It can search history, load older messages, and pull media from conversations you already have. It cannot recover messages the session never received, and it reads nothing until you connect a live WhatsApp session to your account.
Can ChatGPT use the same WhatsApp tools as Claude? Yes. ChatGPT reaches the identical nine tools, just through the manual developer-mode route rather than a one-click in-product connector. The capability is the same; only the setup differs. Claude is the only client with a live in-product connector today, and there is no Gemini path.
