WhatsApp MCP Server

Wabery ships a WhatsApp MCP server so any MCP-capable AI agent, Claude, ChatGPT, Cursor, Codex, or your own, can build on the WhatsApp Business API by describing what you want instead of wiring up the Meta Cloud API yourself. There are two ways to connect: a hosted server with no install, or a local server printed by the Wabery CLI.

Hosted MCP server

The no-install path. Point any MCP client at the hosted server and sign in with OAuth: there is no API key to copy and no CLI to install.

https://mcp.wabery.com

GPT

ChatGPT

Use Wabery from ChatGPT. You sign in with your account, no key to copy.

Open ChatGPT

Step 1 of 13

1
Open ChatGPT in your browser and sign in. You need a paid plan for this: Plus, Pro, Business, or Enterprise.

Open ChatGPT in your browser and sign in. You need a paid plan for this: Plus, Pro, Business, or Enterprise.

Open ChatGPT's Security settings directly. This is the settings page where Developer mode is activated.

Open Security settings

On Security, scroll down, look for Developer mode, and switch it on. A warning pops up: read it and accept.

Open the ChatGPT Plugins page. At the top you will see two tabs, Plugins and Skills. Click Plugins so it is the one selected.

Open Plugins page

Click the round + button on the right. A window titled New Plugin opens.

In the Name box, type Wabery.

Optional: give it the Wabery logo.

Download Wabery logo

Under Connection, leave the switch on Server URL. Paste this address into the box underneath:

Under Authentication, check that the dropdown says OAuth. You can ignore Advanced OAuth settings; Wabery fills that in by itself.

Tick the box that says I understand and want to continue, then click Create.

A Wabery sign-in page opens. Sign in and approve access. There is no key to copy or paste anywhere.

Open a new chat. Click the + button next to the message box, click Developer mode, and switch Wabery on.

Copy the prompt below and send it as your first message.

Client	Setup
Claude	Open Customize → Connectors, click +, then add `https://mcp.wabery.com`.
Claude Code	`claude mcp add --transport http wabery https://mcp.wabery.com`, then `/mcp` to sign in.
Cursor	Customize → MCP → add a server with URL `https://mcp.wabery.com`, then click “Needs login”.
Codex CLI	Add `[mcp_servers.wabery]` with `url = "https://mcp.wabery.com"` and `auth = "oauth"` to `~/.codex/config.toml`, then run `codex mcp login wabery`.

Unlike the local server below, the hosted server has write tools enabled. Your OAuth scopes decide what an agent is allowed to change. Broadcast delivery uses the separate broadcasts:send scope. Every high-risk hosted action, including destructive changes, publishing, function deployment, and broadcast delivery, requests an MCP elicitation that the person must accept in the client; an agent cannot approve its own request. The action fails closed when a hosted client does not support elicitation.

Local MCP server (CLI)

Prefer a local process? Install the Wabery CLI, then print an MCP server config for your agent and drop it into its config file — Claude Code, Codex, and OpenCode are supported:

wabery mcp-config claude --raw
wabery mcp-config codex --raw
wabery mcp-config opencode --raw

Then just describe what you want:

The MCP server exposes the core Wabery primitives an agent needs to build an integration: projects, channels, routing, contacts, registration intents, conversations, message history, outbound messages, inbound media metadata, templates, broadcasts, WhatsApp Flows, hosted functions, submissions, dispatches, automations, webhooks, config-as-code, and Meta Business Agent tools. The local server starts read-only by default; pass --write (or set WABERY_MCP_MODE=write) to allow mutations in a trusted local session.

Hosted function MCP tools let an agent list, create, deploy, test, expose, invoke, tail logs, and delete project-scoped server-side handlers. Function write tools require confirmation tokens in write mode, matching the CLI’s safety model for config apply and flow publish.

For media-aware agents, use wabery_list_conversation_messages to inspect history. Retained inbound WhatsApp files appear as media/assets metadata with short-lived signed URLs; download them promptly and treat expires_at as the retention deadline.

For template and locale-aware workflows, agents should inspect publish_readiness on wabery_get_channel before creating templates. Templates need a dedicated non-sandbox WhatsApp channel, Meta business verification, and a Meta payment method; blockers such as BUSINESS_NOT_VERIFIED or NO_PAYMENT_METHOD mean the template should not be submitted yet. When enrolling contacts, set preferred_language if later flow sends should resolve a locale-specific config key. After wabery_send_message, use wabery_get_message to inspect queued/provider status and Meta failure details. Use wabery_get_template with refresh: true, or wabery_wait_template in scripts, to track template approval.

Broadcast campaigns

Agents can operate the broadcast preparation and delivery workflow with these tools:

Stage	Tools	Required scope
Find an audience	`wabery_list_contacts`, `wabery_list_contact_imports`, `wabery_get_contact_import`	`contacts:read`
Inspect campaigns	`wabery_list_broadcasts`, `wabery_get_broadcast`, `wabery_list_broadcast_recipients`	`broadcasts:read`
Build a draft	`wabery_create_broadcast`, `wabery_prepare_broadcast_audience`, `wabery_duplicate_broadcast`	`broadcasts:write`
Deliver	`wabery_send_broadcast`, `wabery_schedule_broadcast`	`broadcasts:send`
Stop remaining work	`wabery_cancel_broadcast`	`broadcasts:write`

The safe agent sequence is:

Verify there is a dedicated WhatsApp channel and approved template variant for each intended language.
Verify category-specific consent. A phone number or CSV row alone is not consent.
Create a draft, then prepare exactly one audience source: contact_ids, import_id, or filter.
Poll wabery_get_broadcast while status is preparing.
At ready, show the user the total/skipped counts and, when useful, inspect the frozen snapshot with wabery_list_broadcast_recipients.
Ask for explicit confirmation. The hosted MCP server then opens a client-controlled confirmation dialog before send or schedule executes.

Local stdio MCP uses the documented confirmation token fallback in explicit write mode. Canceling stops recipients that have not been dispatched, but messages already accepted by Meta cannot be recalled. Preparing an audience freezes it; duplicating a campaign copies only the template configuration, so the new draft needs a new preparation step.

OAuth authorization codes and refresh tokens are bound to the canonical MCP resource. Redirect URIs must use HTTPS, except HTTP loopback callbacks for installed clients. Refresh tokens rotate on every use; clients must persist the new refresh token returned by the token endpoint. Hosted access tokens are also audience-bound at the API layer: direct replay against the public REST API is rejected because only the MCP service holds the required service proof.

CSV uploads are deliberately handled by the dashboard or CLI rather than passed through an MCP prompt. Run wabery contacts import, poll with wabery contacts import-status, then the agent can find the completed import through wabery_list_contact_imports. For smaller manual audiences, enroll contacts with wabery_enroll_contact.

See WhatsApp broadcasts for request bodies, lifecycle states, frequency safeguards, webhook events, and CLI/SDK examples.