You open WhatsApp to check one thing and come out having skimmed forty threads, replied to two, and forgotten the client who asked you a direct question on Friday afternoon. Nothing in that loop is hard. It is just that nobody reads the pile before you do. Here is the routine that fixes it: three prompts, one triage rubric, and a habit that keeps you approving every outbound message.
What does an AI WhatsApp assistant actually do all day?
An AI WhatsApp assistant reads your existing chats, ranks which threads need a human, and drafts replies you approve before anything sends. Connected to Claude, it runs a read, triage, draft loop on the number you already use.
- Read: pulls the chats and messages inside a time window you name.
- Triage: ranks threads by who is waiting on you, not by recency.
- Draft: writes the reply in your voice and holds it.
- Act: sends or schedules only after you say go.
People search for this as a WhatsApp chief of staff AI, and the name fits. A chief of staff does not make your decisions. They decide what reaches you, in what order, with the context attached. Why that interface is arriving now, and why it is an MCP server rather than another dashboard, is the argument in our piece on WhatsApp going agent-native. This one is the operating routine.
Why the read path matters more than the send path
Your bottleneck on WhatsApp is attention, not typing. You already type fast. What you cannot do is hold forty threads in your head and rank which three deserve the next ten minutes. The best AI assistant for WhatsApp is the one that reads before it writes.
The minutes do not go into composing messages, which take seconds. They go into the scan: opening a thread to remember what someone wanted, re-reading a chain to check whether you replied, working out whether a group chat with 60 unread messages contains anything addressed to you.
Reading is also the low-consequence half. Nobody receives anything, nobody gets messaged twice, and a bad output costs you a re-prompt instead of an apology. Be aggressive on the read path and conservative on the send path, which is the reverse of how automation usually gets pitched. For the wider surface a WhatsApp AI agent covers, that piece maps it.
What you need to read WhatsApp with Claude
Three things: your own WhatsApp number connected over WhatsApp Web or a managed gateway, an MCP connection between Claude and that number, and about two minutes. This runs on your personal number and your real chat history. It is not the Meta Cloud API and not the WhatsApp Business API.
That distinction is load-bearing. Meta's own docs state that a registered business phone number cannot be used with WhatsApp Messenger, and that numbers already in use with WhatsApp cannot be registered unless they are deleted first. The Cloud API sends business messages from a separate identity, billed per delivered template message since July 1, 2025. It cannot brief you on your personal conversations, because those conversations do not live on it.
Three ways to wire up the connection that can:
- Remote connector (no key). Add
https://api.blueticks.co/mcpas a custom connector in Claude and approve access with a login. Note the path:/mcp, not/mcp/sse. Anthropic documents custom connectors over remote MCP as available on Claude, Cowork, and Claude Desktop across the Free, Pro, Max, Team, and Enterprise plans, authorized through OAuth rather than a pasted key. - npm package (API key). Run
npx -y @blueticks/mcpwith an API key for stdio clients like Claude Code and Cursor. - REST API. Writing your own code? The
/v1API covers the same actions, documented at dev.blueticks.co.
The click-by-click version, including where each client stores its config, is in the Claude WhatsApp connection guide.
How to run a morning WhatsApp brief with Claude
Ask for a narrow window and a capped list, never a summary of everything. A good brief names a start time, restricts scope to threads where someone is waiting on you, caps the length, and asks for one line of context per thread. You get a shortlist you can clear in ten minutes.
Start with these three.
"Catch me up on WhatsApp since Friday 5pm. List only threads where someone is waiting on me, newest first, max 10, one line of context each."
"Of my one-to-one chats from the last 24 hours, which mention a date, a price, or a decision? Skip group chats entirely."
"Read the last 30 messages in my thread with the Riverbend Supply account and tell me, in three bullets, what I owe them."
Aim for this output shape. The content below is invented for illustration and is not anyone's real WhatsApp:
Waiting on you since Friday (4)
- Supplier contact, 3 days: asked you to confirm the Thursday delivery slot.
- New enquiry, 2 days: asked what the setup fee is, never got an answer.
- Existing client, 1 day: wants to move the Wednesday call, proposed two times.
- Accountant, 6 hours: chasing one missing receipt. Handled or no action: 31 threads.
Who, what, and how long they waited. No recap of threads that resolved themselves. Four levers get you there:
- Time window. "Since Friday 5pm" beats "recently" every time.
- Scope. Say one-to-one or groups, unread or all. Groups are where the noise lives, so exclude them by default.
- A cap. "Max 10" forces ranking. Without it you get timestamp order, which is the pile you were escaping.
- One line of context. Enough to decide, not enough to re-read.
The failure mode is asking for too much. The assistant works from the chats and messages it can actually read, so a two-week window across every group costs time and dilutes the ranking. A tight window run twice a day beats one heroic sweep. The read tools behind it are listed in our WhatsApp MCP primer.
How to get Claude to triage what needs a human today
Give it a fixed rubric instead of asking what is important. Buckets plus your own keyword rules produce a repeatable ranking, and the deliverable is a short list of things to do today. Paste the same rubric every morning so the output stays comparable day to day.
Edit the names and keywords to yours, then keep this one keystroke away:
"Sort what you found into exactly four buckets: NEEDS ME TODAY, WAITING ON THEM, I PROMISED SOMETHING, NOISE. Rules: anything mentioning invoice, refund, reschedule, or cancel goes to NEEDS ME TODAY. Anything from Riverbend Supply, Acme Dental, or K. Alvarez goes to NEEDS ME TODAY regardless of content. Group chats go to NOISE unless someone used my name or asked a direct question. If my last message in a thread ended in a question, it is WAITING ON THEM. Then show me NEEDS ME TODAY and I PROMISED SOMETHING only, max 7 lines each, and give me the count of the other two buckets without listing them."
Two things it does on purpose. It encodes your priorities as named accounts and literal keywords, which a model handles reliably, rather than as a vibe. And it makes silence visible: WAITING ON THEM is the bucket you never build by scrolling, because a thread you are waiting on looks identical to a finished one. The output cap matters too. A triage list you have to triage is not triage.
Then add the end-of-day sweep:
"Look at every message I sent today. List anything I committed to that I have not actually done or scheduled yet."
That one stings the first few times, and it is the highest-value prompt here, because the promises you forget are the ones nobody chases until they get expensive. Send-side counterparts, paced follow-ups and campaigns, live in automating WhatsApp with AI.
Ask Claude for your first morning WhatsApp brief. Connect your own number in a click at blueticks.co/signup. No Meta API, no per-message fees, no local server to keep alive.
How to make Claude draft replies in your voice without letting it send
Separate drafting from sending, every time. Ask for a draft, tell it to read the thread for tone, review the text, then say send. Or schedule it for business hours instead of firing at 1am. The approval step is not friction to optimize away. It is why this workflow is safe to run on your real number.
The protocol is built on that assumption. The Model Context Protocol specification says that for trust and safety "there SHOULD always be a human in the loop with the ability to deny tool invocations," and that clients should show tool inputs to the user before calling the server. Anthropic's connector guidance is equally direct: review tool approval requests carefully and only click "Allow always" for a server and tool you trust to run unsupervised. Read tools are a fair candidate for that. A send tool is not, and I would leave it on per-call approval permanently.
Voice is an instruction problem, and the instruction that works is specific and negative:
"Draft a reply to the client who wants to move Wednesday. Match how I wrote earlier in that same thread, two sentences maximum, no emoji, no exclamation marks, and do not apologize. Show it to me, do not send it."
"Write it in my voice" gets you a generic professional register. "Match how I wrote earlier in that same thread" gets you your actual one, because your previous messages are right there. When a draft comes back wrong, correct it in one line the way you would correct a capable new hire: "too formal, drop the second sentence."
Then pick your moment. Approving a message at 11pm does not mean it should leave at 11pm. Say "schedule that for 8:30am tomorrow" and it queues instead of landing while your client is asleep. Timing details are in scheduling WhatsApp messages from Claude.
Which chief-of-staff jobs to hand over first
Start where the cost of missing something is high and the cost of a wrong answer is low: unanswered leads, then your own broken promises, then group noise, then confirmations, then the weekly recap. Each is one prompt with an output you can judge at a glance.
| Job | The prompt | What good output looks like |
|---|---|---|
| 1. Unanswered-lead sweep | "Find chats from the last 14 days where someone asked me a question and my last message came before theirs. Enquiries first." | A short list of names, each with the question that never got an answer. |
| 2. Promises you did not keep | "What did I say I would do in the last week that has not happened yet?" | Your own commitments, quoted back with the thread and the date. |
| 3. Group-noise filtering | "From my group chats in the last 24 hours, show only messages that name me or ask a direct question." | Two or three lines, or nothing at all on a quiet day. |
| 4. Supplier confirmations | "Which suppliers confirmed something this week, and which are still silent on a date I asked about?" | A confirmed column and a chasing column, no prose. |
| 5. Weekly recap | "Summarize my WhatsApp week: what closed, what is still open, who is waiting on me going into Monday." | One paragraph plus an open-items list for Monday. |
Job 1 usually pays for the setup on its first run, because there is almost always a lead in there who asked a real question and never heard back. Job 2 changes your behavior. Job 3 you will keep forever, because group chats are where useful messages go to die. Turning a found lead into an actual sequence is covered in our guide to WhatsApp follow-up reminders.
What an AI WhatsApp assistant cannot do, and where the line is
It is not an official Meta product, it does not use the Meta Cloud API or the WhatsApp Business API, and no tool can promise your number will never be banned. Reading your own chats is low-risk. Sending through automation on your own number carries the same exposure as any WhatsApp Web tool, and consent for every message is yours.
Be precise about that exposure. WhatsApp's Terms of Service prohibit conduct that involves "sending illegal or impermissible communications such as bulk messaging, auto-messaging, auto-dialing, and the like," and prohibit non-personal use of the service unless authorized (WhatsApp Terms of Service). Any tool driving a personal number sits inside that reality. What moves your risk is behavior: volume, cold first contact, and how many recipients block or report you. Not which vendor you picked.
The second limit is judgement. It ranks by the rules you gave it, so it will confidently drop a thread into NOISE that mattered for reasons you never wrote down. It does not know that the short message from your biggest customer was a warning sign, or that a supplier's "no problem" has meant a delay three times running. It drafts, you decide.
The alternatives deserve credit, because they came first and they are good software. lharries/whatsapp-mcp pairs a Go bridge on whatsmeow with a Python MCP server and keeps your history in a local SQLite file. WAHA is a self-hosted gateway that can expose an MCP endpoint. Composio routes WhatsApp tools through its connector layer. If you want local control and your data on your own disk, those are the better choice, and our comparison of WhatsApp MCP servers lays out the trade-offs properly.
What a chief-of-staff setup needs to be reliable at 7am
It needs to be awake before you are. A morning brief is worthless if the connection died overnight, and a local bridge dies quietly: your laptop sleeps, the process drops, and the 7am digest does not happen. Nothing errors. You find out at 11am when a client follows up.
That is the gap Blueticks fills. It is a hosted remote OAuth MCP at https://api.blueticks.co/mcp plus a /v1 REST API, both on a managed 24/7 WhatsApp engine, so the session stays up without you babysitting it. The same connection covers the write path once you approve one: send now, schedule for 8:30am, or run a paced campaign. For sends that fire while your machine is off, that is the offline gateway, and the same engine serves the API at dev.blueticks.co if you would rather script the brief than ask for it.
The claim, stated as narrowly as it deserves: a first hosted, no-code, product-grade agent-native WhatsApp interface. Not the first WhatsApp MCP. The open-source servers got there before us. What is new is that the agent is the primary user and there is no bridge for you to keep alive.
Ask Claude for tomorrow's morning brief instead of scrolling for it. Connect your own number at blueticks.co/signup: no Meta API, no per-message fees, no local server to keep alive.
Frequently asked questions
Can AI read my WhatsApp messages? Yes, through your own connection. Once your number is linked and Claude is connected over MCP, it can list your chats, read the messages inside a thread, and search history, which is what makes digests and triage possible. It reads your real conversations because it drives your real session.
Can Claude reply to WhatsApp for me? It can draft and it can send, but keep those separate. Ask for a draft, review it, then approve the send or schedule it for business hours. The MCP specification says there should always be a human able to deny a tool call, and a send tool is exactly the kind to approve individually.
Is this the WhatsApp Business API? No. It runs on your own number over WhatsApp Web or a managed gateway. It is not the Meta Cloud API or the WhatsApp Business API, which sends from a separate registered business number and bills per delivered template message. A Cloud API number cannot be used with regular WhatsApp and cannot read your existing personal chats.
Will an AI assistant get my WhatsApp number banned? There is no guarantee either way, and anyone offering one is overclaiming. Reading your own chats is low-risk. Sending through automation on your own number carries the same exposure as any WhatsApp Web tool, since WhatsApp's terms prohibit bulk and auto-messaging. Message people who expect to hear from you.
Do I need to write code?
No. The connector path is a URL and a login: add https://api.blueticks.co/mcp as a custom connector in Claude and approve access. Everything in this article is typed in plain language. Developers who want the same actions from their own code can call the /v1 REST API instead, documented at dev.blueticks.co.
Does it work when my laptop is off? Only if the session is hosted somewhere that stays awake. A self-hosted local bridge stops when your machine sleeps, so a 7am brief silently does not run. Blueticks runs the WhatsApp session on a managed engine, so the connection and any scheduled sends survive your laptop being shut.
