Connecting Claude to your WhatsApp account

You can connect Claude directly to your own account, so it can read your conversations, inspect the automations you already have, and build new ones from a plain-English description. It runs over MCP — an open standard for connecting AI assistants to tools — and it is set up with a single command.

The reason this is interesting is not novelty. It is that the hard part of automation was never drawing boxes on a canvas; it was knowing what to build. An assistant that can read what your customers actually ask is working from evidence rather than from your recollection.

What it can do

Twenty-three tools, grouped by job:

Area	What Claude can do
Learn from real chats	List conversations, read one, search messages, and group recurring themes
Inspect what you have	List automations, read what a flow actually does, and see who changed it and when
Build and edit	Create an automation, update an existing one, run a quality check, publish
Set up prerequisites	Create tags and contact attributes, and repair tag steps given a name instead of an id
Knowledge base	Suggest questions customers asked that the bot could not answer, and add approved answers
Reporting	Analytics overview and per-agent performance

Connecting it

1. Go to Settings → Integrations and find the Claude card. It is admin-only.
2. Generate your access token. It starts with cfm_.
3. Copy the connect command Settings gives you and run it in Claude Code.
4. Ask Claude something simple to confirm it worked — "what are my customers asking about most?" is a good first test.

What it is genuinely good at

Finding what to automate. Ask it what customers ask most and it groups your real conversations into themes. That is a better starting point than any list of use cases, because it reflects your customers rather than a generic store.

Filling the knowledge base. It can surface the questions your bot could not answer, ranked by how often they came up, alongside what your agents actually replied. You approve or correct an answer and it gets added — which turns a blank knowledge base into a useful one in an afternoon.

Explaining a flow you inherited. "What does this automation actually do?" is a real question when a rule has been edited by three people, and the version history says who changed what and when.

Building the first draft. Describing a flow in words and getting a working draft is faster than dragging steps, especially for a branching flow where the fiddly part is wiring every path.

Nothing goes live without you

This is the part that makes it safe to try. An automation Claude creates is saved as a disabled draft — it does not run and it cannot message anyone. You open it in the builder, read every step, and publish it yourself.

Publishing through Claude is possible too, but it shows you what will go live first. Either way, the same checks apply as any other publish: a flow with an unselected template is blocked rather than going live and failing silently.

The privacy position, stated plainly

Your chat transcripts leave the platform. When Claude reads a conversation, that content goes to your MCP client and your model provider — that is inherent to how the feature works, not a detail we can engineer away.

Three consequences worth weighing:

• The token is admin-only, because it opens every conversation in the organisation.
• Real customer phone numbers, addresses and order details are in those transcripts.
• If your market has strict data-protection rules, decide whether this is appropriate before generating a token rather than afterwards.

You can rotate or revoke the token from the same Settings card at any time.

Good first requests

1. "What are my customers asking about most this month?"
2. "Which questions has the bot failed to answer?"
3. "Read my order confirmation automation and tell me what it does."
4. "Build me a flow that catches return requests, asks for the order number and reason, and hands it to a person."
5. "Check my automations for problems before I publish."

That last one is worth knowing about. There is a check that looks for the faults publishing would have caught — a tag step pointing at a name rather than a tag, a variable nothing sets, a branch wired to nowhere.

Where it does not help

• It cannot see your stock or your prices unless they are in a conversation or a sheet it can read.
• It does not change Meta's rules. A flow it builds still needs approved templates to reach a cold chat.
• It cannot judge your business. It can tell you what customers ask; whether answering that automatically is a good idea is still your call.

Why this is unusual

Most platforms in this category offer an AI that writes message copy. This is a different thing: an assistant with read access to your actual operation and the ability to build in it, with a human approval step in front of anything that goes live.

If you have been putting off building automations because the blank canvas is the obstacle, this removes the obstacle. Start with building your first automation for what a good flow looks like, then ask Claude to build the next one.

What it cannot change about Meta's rules

An assistant that builds flows does not get different treatment from Meta. Every automation Claude writes lands under the same constraints as one you draw yourself:

• Reaching a cold chat needs an approved template. The 24-hour customer service window opens when a customer messages or calls you and resets each time, per Meta's sending messages documentation.
• Your messaging tier still caps outbound reach — it starts at 250 unique recipients per rolling 24 hours and runs 250 → 2,000 → 10,000 → 100,000 → unlimited, per Meta's messaging limits documentation.
• Template ceilings still apply — 250 per WhatsApp Business Account unverified, up to 6,000 once verified, per Meta's template documentation.

What changes is how quickly you get from an idea to a correct draft, not what the draft is allowed to do.

How it fits with building by hand

Most merchants end up doing both, and the division settles quickly. Claude is faster at the first draft of a branching flow and at anything requiring it to read across many conversations. The canvas is better for a small edit you can see, and for understanding what a flow does at a glance.

Because everything arrives as a draft and the version history records who changed what, moving between the two is safe. If an AI edit makes a flow worse, the previous version is still there.

A sensible first session

1. Ask what customers ask about most, and read a couple of the real conversations it points at.
2. Ask it to list your existing automations and explain one.
3. Ask it to build the flow for your most common question.
4. Open that draft in the builder and read every step.
5. Publish it yourself once you are satisfied.

Step four matters. The draft is a starting point you are responsible for, and the review is where you catch anything that does not match how your business actually works. Testing protocol is in testing an automation safely.

Where it fits alongside the builder's own AI

The automation builder already has an AI that can generate or edit a flow from a description, working inside a single automation. The Claude connection is broader: it can read your conversations to work out what to build, look across every automation you have, and set up the tags and contact fields a flow depends on before building it.

In practice the builder's AI is the faster route for "change this step"; Claude is the better route for "look at my customers and tell me what I am missing". Both leave you with something to review rather than something already live.

A note on cost

The connection itself is part of your plan; what it consumes is your own AI usage on whichever client you connect from. Reading a large number of conversations costs more than reading a few, which is why response payloads are capped — one call cannot flood the client with your entire message history.

Practically that means asking focused questions gets better answers and costs less: "what did customers ask about returns this month" beats "read everything". The same discipline applies to the flows it builds — a focused brief produces a flow you can read, and every automation step, explained is the reference for checking one.

Revoking access

Rotate or revoke the token from the same Settings card. Worth doing if the person who set it up leaves, if you connected from a machine you no longer control, or simply on a schedule — the token opens every conversation in the organisation, so it deserves the same treatment as any other admin credential.