Use Met from Claude in 3 steps
Connect your Met account to Claude, ChatGPT or Cursor without writing code or creating an API key, and ask it in your own words to answer, look up and build.
Your AI assistant —Claude, ChatGPT, Cursor or any other that speaks MCP— can work inside your Met account. You write to it the way you would write to someone on your team («which chats haven't I answered?», «create an orders collection») and it does it in Met for you.
You don't need to code or create an API key: you paste an address, sign in to Met and approve the permissions.
Step 1 · Add Met as a connector
The Met server address is always the same:
https://api.met.meteor.com.co/api/v1/mcp
Where you paste it depends on your assistant:
- Claude (web or desktop app): in Settings → Connectors, add a custom connector, name it
Metand paste the address. If your Claude account belongs to a company, whoever manages it may have to add it first for the whole team. - ChatGPT: in Settings → Apps & Connectors, create a connector with the same address. Depending on your plan, you first have to turn on developer mode in the advanced settings.
- Cursor: in Settings → MCP, add a new server, or paste this into
.cursor/mcp.json:
json { "mcpServers": { "met": { "url": "https://api.met.meteor.com.co/api/v1/mcp" } } }
- Claude Code: in the terminal,
claude mcp add --transport http met https://api.met.meteor.com.co/api/v1/mcp, then/mcpto sign in.
Step 2 · Sign in and approve the permissions
When you connect, your assistant opens a Met window. Sign in with your usual user and review the list of what the assistant will be able to do: read contacts, send messages, create flows, and so on. If you agree, tap Authorize.
- The assistant works on the workspace you have open in your session. If you manage several, switch to the one you want to connect first.
- Met never hands your password to the assistant. The session renews itself, so you don't have to sign in again every day.
- You can remove its access at any time in Settings → Developers → Authorized apps. The cut is immediate.
Step 3 · Ask for what you need
Write in your own words. You don't need to know what each tool is called: the assistant picks the right one and, if there is no specific tool for what you ask, it looks up the Met operation that does it and uses it with the same permissions you approved.
For day-to-day work
- «Which conversations haven't I answered since yesterday?»
- «Find Laura Gómez in my contacts and tell me what she last wrote to us.»
- «Reply to Laura that her order ships tomorrow morning.»
- «Schedule a reminder for Carlos on Friday at 10 a.m. about the quote.»
- «Hand Andrés's chat to someone on the team and tag it as *Complaint*.»
- «How much Energy do I have left, and which Met is spending the most?»
For building
- «Create a Met that answers questions about my product catalog, in a friendly tone, and hands the chat to a person when someone asks about warranties.»
- «Build a flow that, when a new lead comes in, tags it as *New* and lets me know.»
- «Create an *Orders* collection with the fields customer, product, amount, delivery date and status.»
- «Schedule an automation that sends me the pending orders summary every Monday at 8 a.m.»
- «Show me the templates I can install and tell me what each one would create.»
A new flow stays as a draft: the assistant builds it, can test a single step and asks you before publishing it. Until it is published, it doesn't trigger on its own.
What it asks you before doing
Some actions Met won't allow without your explicit confirmation: the assistant has to ask you first, and only goes ahead if you say yes.
- Deleting anything: a flow, an automation, a collection, a field, a record.
- Publishing a flow: from that moment its triggers use the new version.
- Replacing something that already exists: overwriting a flow's draft or an integration's credentials.
- Doing a lot at once: installing a template, importing contacts or records in bulk, tagging many contacts, creating a broadcast.
- Running a scheduled automation right now, outside its schedule.
What costs money
Every tool that spends says so, and the assistant warns you before using it:
- WhatsApp templates and broadcasts: every message has a cost with Meta and reaches a real person. The assistant confirms the template and the recipient with you, and can show you the message exactly as it will go out before sending it.
- Energy: running a Met, testing a flow step or running an automation spend Energy, just like when they happen on their own.
Creating or editing Mets, flows, collections and records doesn't spend Energy. Neither does looking things up.
On top of that, your assistant can ask for your permission every time it is about to use a Met tool. You set that up in the assistant's own settings.
There is one thing the assistant won't do even if you ask: talk over your team. If a person is handling a chat —autopilot is off or paused, or the contact asked to talk to a human—, the message isn't sent and the assistant tells you so.
What you can't do from the assistant
- Upload files: images, PDFs or documents are uploaded from the Met app. The assistant can still find the files you already have and give you their name and link.
- See secrets: API keys, integration credentials or signing secrets never reach the assistant. It can store a credential you give it, but not read it back afterwards.
- Buy or change plans: topping up Energy, adding add-ons or changing plans happens in the app. The assistant can tell you how much you have left and where it's going.
If your plan expired, the assistant can keep looking things up (contacts, conversations, data), but it can't change anything until the account is up to date. It tells you so instead of trying over and over.
If something doesn't work
- What you ask for doesn't show up: every permission you didn't approve leaves its tools out. Disconnect Met in your assistant, connect it again and approve what's missing.
- It works on the wrong workspace: remove its access in Settings → Developers → Authorized apps, open Met in the right workspace and connect again.
- You want the technical detail —connecting with an API key, the tools by area, how each one behaves—: it's in MCP server.