What can the Klaviyo MCP server actually do?

What is the Klaviyo MCP server?

MCP, the Model Context Protocol, is the standard that lets AI assistants use external tools. Klaviyo’s MCP server is, in its own help centre’s words, “a wrapper on top of Klaviyo’s APIs”: it gives an AI client a catalogue of Klaviyo actions it can call on your behalf, with your permission, against your account data.

The practical effect is that Claude, ChatGPT or Cursor can answer questions like “which flows drove the most revenue from returning customers last quarter” by actually pulling your data, rather than guessing. Everboost uses it in production client work, which is the vantage point for this guide, including for its limits.

The timeline matters because a lot of coverage is stale:

Date	What shipped
June 2025	Local server first published (runs on your machine, private API key)
August 2025	Beta announced, then general availability with a remote hosted server
January 2026	Listed connector in Claude’s directory; Klaviyo app in ChatGPT
May 2026	Anthropic partnership expansion: attributed reporting data, interactive MCP apps in Claude
July 2026	Klaviyo publishes dedicated agency guidance
September 2026	K:BOS: “Klaviyo Headless”, 260+ MCP tools and 490+ APIs callable from any AI system
October 2026	Live connections expose flow, segment and form creation (drafts) and confirmed campaign sends

Articles describing a beta you install with npm are describing mid-2025. As of July 2026 the normal path is the hosted server at mcp.klaviyo.com/mcp with OAuth and nothing to install.

How do you connect it?

Three requirements catch people out. You need an Owner, Admin or Manager role in Klaviyo. On Claude, the connector needs a paid plan (Pro, Max, Team or Enterprise). And the connection binds to whichever Klaviyo account you are logged into when you authorise it, which matters enormously if you touch more than one account; more on that below.

Claude users add Klaviyo from the connector directory in a few clicks. ChatGPT users add the listed Klaviyo app. Cursor and VS Code connect to the same remote URL. The local server still exists for developers who want to run it against a private API key.

One configuration detail worth knowing exists: the remote server accepts query parameters, including a full read-only mode, but Klaviyo notes these cannot be set on the listed Claude connector or ChatGPT app. If you want read-only, you add Klaviyo as a custom connector pointing at the same URL with read-only=true.

What can it actually do, and what can it not?

This is where most coverage stops at the marketing copy, so here is the capability map as it stands in October 2026, from Klaviyo’s documentation checked against a live production connection. It has moved a long way since July.

It reads almost everything. Campaign, flow and form reports, metrics and attributed aggregates, segments and lists, profiles, catalogues, templates, deliverability-relevant account data. Reporting and audit questions are its natural territory.

It now writes almost everything that matters. In July it could create campaign drafts, templates and profiles but could not touch flows or segments. As of October 2026, Everboost’s production connection also exposes tools to create flows from a full definition (trigger, delays, conditional splits, email, SMS, WhatsApp and webhook actions, new actions in draft), create segments, create forms (as drafts), manage Customer Agent skills and knowledge, and send campaigns. The send tool refuses to run until the user explicitly confirms.

The ChatGPT app is more limited by design. Klaviyo’s documentation says it cannot send campaigns, build or edit flows, create SMS campaigns or modify segments and lists, so check which surface you are on before assuming a capability.

The documentation trails the live server. At K:BOS in September 2026 Klaviyo’s press release counted “260+ MCP tools and capabilities” while its community explainer said “more than 400 tools”. The tool list on your own connection is the only reliable answer. What changed at K:BOS, and what else Klaviyo announced, is in our K:BOS 2026 breakdown.

The shape of this map has changed: flows and audience definitions, the highest-consequence surfaces, are now writable by an AI tool. Klaviyo has kept the guardrails at the edges (new flow actions start as drafts, sends need explicit confirmation), which makes read-only mode the right default for every connection that does not need to build.

Why don’t the numbers match your dashboard?

The single most common complaint from real users, visible across Reddit and Klaviyo’s own community, is that the MCP returns revenue figures that disagree with the dashboard. This is not a hallucination problem. It is an attribution problem, and it has a fix.

Your dashboard shows last-touch attributed revenue inside Klaviyo’s attribution windows, typically five days for clicks and one for opens. An AI asked casually for “campaign revenue” will often query raw Placed Order events in the period instead, which counts every order regardless of what drove it. Two different questions, two different numbers, neither wrong.

Since May 2026, the server’s Query Metric Aggregates tool exposes “the same underlying data that powers Klaviyo’s in-app Metric Reporting”, so the matching number is available if the AI queries it. The working habits Everboost applies on client accounts:

1. Ask for attributed metrics explicitly, naming the attribution model when it matters.
2. Spot-check any number that will be repeated to a stakeholder against the dashboard before it travels.
3. Treat matching numbers as a property of the question you asked, not the tool.

Can an agency run it across multiple client accounts?

Here is the constraint Everboost hit in production in May 2026, before Klaviyo documented it: one connection binds to one Klaviyo account. Switching means disconnecting and reconnecting, and the account is determined by whoever is logged in at OAuth time.

Klaviyo’s agency guide, published July 2026, formalises the workaround: a separately named custom connector per client and a Claude project per client instructed to use the right one. That works for strategist work, one client at a time, and it is genuinely useful there: audit questions that used to mean an hour of clicking come back in seconds.

What it does not fit is multi-client automation, anything that needs to run across ten accounts on a schedule. For that, Everboost bypasses the MCP entirely and uses the Klaviyo REST API with per-client private keys, which is the architecture Klaviyo’s own rate-limited, key-scoped API model is built for. Use the MCP as a strategist’s instrument, not an integration layer.

Is it safe to let an AI touch your Klaviyo account?

Klaviyo has done more here than most realise, and the controls are worth using.

Read-only mode disables every write tool: the right default for analysis work and for any connection a client can see. The user-generated-content setting disables tools that read profile data fields, which is Klaviyo’s defence against prompt injection, the scenario where malicious instructions hidden in a customer’s profile get read and acted on by your AI. Klaviyo’s own docs say to “carefully review each tool call”. That is the correct paranoia.

Two trust-model warnings for anyone searching “Klaviyo MCP” and landing on GitHub. Unofficial community servers typically want your private API key in a plain-text file, and third-party MCP platforms route your credentials through their infrastructure. The official server never hands your key to anyone. For client accounts there is no good reason to use anything else.

And if your team uses ChatGPT on consumer plans, read Klaviyo’s FAQ closely: OpenAI may train on connector data from Pro and Plus plans by default, but not from Enterprise, Team or Edu plans. For client data that is a plan-selection decision, not a footnote.

What is it genuinely good for?

After months of production use across client accounts, Everboost’s honest scorecard:

Excellent: reporting speed, account audits, campaign drafting, structured questions with a checkable answer. The audit use case alone justifies connecting it.

Adequate with supervision: template work, segment and flow drafts (built as drafts, then checked by a human in the editor), performance summaries for internal use.

Not yet: complex multi-step analysis. Cohort repurchase behaviour, product-journey questions and category-level flow analysis are where users report wrong answers and heavy token consumption, and our experience matches. The tool is a fast junior analyst, not a retention strategist: it can fetch and assemble, but deciding what matters, and noticing when an answer is plausible and wrong, still sits with whoever is accountable for the number.

That last distinction is the one to hold onto. The MCP server collapses the distance between a question and Klaviyo’s data. It does nothing to close the distance between data and a good decision, and a brand’s retention problems live almost entirely in the second gap.