Meta Ships WhatsApp Business MCP for Claude, Cursor, Codex
Meta's new MCP server lets coding agents create WhatsApp Business accounts, verify numbers and write templates. What to test before you hand it the keys.
Meta shipped a WhatsApp Business Tools MCP server on Tuesday, letting an AI coding agent do account setup that previously spanned four separate tools.
TechCrunch reports the server "directly connects an AI coding agent like Claude, Cursor, Codex, or ChatGPT to the WhatsApp Business Platform." Before it, developers had to move between the Developer Console, Meta's Business Manager, the API reference, and their editor to get a business onto WhatsApp messaging.
What the agent can actually do
Per Meta's description, the agent handles the busywork of setup: creating the WhatsApp Business account, adding and verifying the phone number, registering it for access to the Cloud API, and checking the business' Terms of Service. Beyond onboarding, it can draft a messaging template from a plain description, edit an existing template, test messages and webhooks, and monitor things that previously failed quietly — Terms of Service status, payment method, and Business Verification.
Meta also points to a second server, Meta Social Technologies MCP, for discovering API endpoints, searching documentation, and troubleshooting errors during setup.
Meta is joining a crowded list
This is not a first mover play. TechCrunch lists PayPal, Stripe, GitHub, Notion, Slack, Salesforce, Atlassian, X, Google and Microsoft as already offering MCP servers for agents to reach their services. For Meta specifically, it extends a lineup that previously covered ads management and app configuration monitoring into business onboarding.
The practical read: if your stack already runs an agent against Stripe or GitHub MCP servers, this slots into the same pattern, with the same operational questions — what the agent is allowed to create, and who reviews it.
The variable that decides whether it works
Whether an MCP server is useful depends less on the vendor than on how its tools are described, and there is fresh data on that. A write-up published the same day by Vinny Barreca reports on Toolmetry, described as "an open-source project that measures how well agents use MCP tools," which had an LLM rewrite existing MCP server tool descriptions and nothing else.
The reported before-and-after, by the author's own account:
| Server | Before | After |
|---|---|---|
| SQLite | 34.0% | 100% |
| Memory | 61.8% | 96.4% |
| Git | 75.0% | 96.7% |
The author puts total API spend for the run at $4 and says the model and the harness never changed. Treat the exact figures as one project's self-reported experiment, not an independent benchmark — but the three failure modes it names are recognisable: two tools with overlapping descriptions so the agent picks the wrong one, descriptions that imply a prerequisite call the agent then makes every time, and parameter names left in the text after the API moved on.
How to evaluate it this week
If you are considering handing WhatsApp onboarding to an agent, the useful test is narrow. Run the setup flow once on a throwaway business account and log every tool call the agent makes — ritual repeat calls and wrong-tool picks show up immediately in that log. Check whether verification and Terms of Service steps report failure loudly or silently, since those are exactly the items Meta says the server now monitors. And keep template creation behind human review at first: a template is customer-facing text, and an agent writing it from a one-line description is the highest-variance part of this whole surface.
Worth watching next: whether Meta publishes scoped permissions for the server, and whether any of these vendor MCP servers start shipping measured tool-use success rates instead of feature lists.
More from DangMua