K:BOS 2026: What Klaviyo going headless means for you

Hey community,

If you only saw the headline slide from K:BOS 2026, you might think Klaviyo launched a chatbot. It didn't. Here’s a deep dive on what actually was announced, why it matters, and how to get started.

In a sentence: Klaviyo opened more of its data and supported actions to Claude, ChatGPT, and other AI tools, so you can work with your Klaviyo account from places you already use, without opening the Klaviyo UI for every question or task.

Yes, that sounds technical, but the customer impact is pretty practical. My biggest impression with all the product updates that were announced echoed that sentiment – things felt practical and more quality improvements versus shiny net new product. And that’s a good thing and likely even more impactful for the community. With the headless part of things, the idea is you get answers faster, automate more of the repetitive work, and bring Klaviyo into the same AI workflows you already use. You also have new decisions to make about permissions, privacy, approvals, and which tasks you actually want an agent to perform.

Those are already the questions customers are raising in the Community:

• How do I limit an AI tool's access to sensitive profile data?
• What does read-only really protect?
• What can MCP do today, and where does it stop?
• Can an agent preserve my templates, flows, and brand context?
• How should I set this up if I manage several accounts?

The most useful way to understand this launch is not as a list of features. It is as a change in how you can use Klaviyo and a new set of choices about how much control you want to give an AI tool.

I watched the K:BOS keynote so you don't have to sit through all 47 minutes. Here's what was announced, what is available now, and where I would start.

What does “headless” mean for a Klaviyo customer?

Headless means you can use Klaviyo's data and capabilities without starting in the Klaviyo interface.

For example, you could ask Claude to:

• Summarize campaign or flow performance
• Identify an audience that is changing behavior
• Answer a question about customer or event data
• Draft a campaign brief based on recent performance
• Recommend what needs attention
• Take a supported action after you approve it

You can also build your own workflow using the APIs or MCP. A team could create a reporting dashboard in another tool, send a weekly performance summary to Slack, or connect Klaviyo to a custom agent.

The important qualifier is supported. What an agent can see or do depends on the connector, your role, the scopes you grant, and which tools are available. Headless access does not mean every AI tool can do everything Klaviyo can do.

What is the SQL preview?

Klaviyo is making it possible to ask data questions in plain language through Composer or MCP and receive both an answer and the SQL query used to produce it.

The initial preview covers areas such as profiles, events, segments, lists, campaigns, and catalogs. Forms, flows, and custom objects are still being added.

Klaviyo also introduced a semantic layer. That matters because business terms are rarely as simple as they sound. “Revenue,” “best customer,” and “high-value segment” can mean different things depending on your definitions, attribution windows, and exclusions.

The goal is for Klaviyo and the agents connected to it to use a more consistent understanding of your data. More customization for your own metrics and definitions is still coming.

When you ask an agent for a number, do not just ask whether the answer is “right.” Ask which data source, metric definition, and attribution window it used. An event-level total and a dashboard metric may be answering different questions.

What is changing with Personalization?

Marketing Analytics is becoming Personalization. This is more than a navigation rename, but it is also not a completely new product replacing everything you use today.

Klaviyo is bringing existing personalization capabilities into a broader decisioning layer. Specialized models can help decide:

• Who to reach
• What to show them
• When to reach them
• Which channel to use

The named capabilities include Audience Optimization, 1:1 Content, Product Recommendations, Discount Decisioning, Personalized Send Time, and Channel Affinity.

The practical change is that these capabilities are being positioned as connected parts of a larger system rather than isolated tools. You still set the guardrails and define what you want to optimize for. The models work inside those constraints, and holdout testing is intended to measure whether the personalization is actually creating lift.

Klaviyo says the platform processes more than 4 billion signals a day across billions of profiles. Because current materials use different profile-count figures, I would not rely on one exact number without checking the latest product messaging.

What is new in Composer?

Composer is moving from a tool you prompt to a tool that can keep work moving in the background.

Depending on your account and rollout status, Composer can:

• Run scheduled or recurring tasks
• Save skills and connectors
• Analyze performance and explain what changed
• Create or iterate on segments
• Return citation-backed answers
• Turn supported recommendations into one-click edits

The important part is the approval step. Composer can prepare or apply supported edits inside the workflow, but you review and approve the work before it goes live. It does not independently send or publish campaigns without sign-off.

There was also a preview of a Composer iOS app, which is intended to bring questions and notifications to wherever you are working.

What is new in Customer Agent?

Customer Agent now supports automatic language detection and response across supported channels, with 100+ languages and Shopify Markets localization in the K:BOS announcement.

Klaviyo also previewed auto-training and profile enrichment. In practical terms, you can refine how the agent behaves in plain language, and shopper preferences or intent can be written back to the customer profile.

That is different from using customer conversations to train a general-purpose large language model. Klaviyo's help documentation says customer data is not used to train large language models.

Voice was also previewed as a future channel. Availability and pricing are still being clarified, so treat it as an early preview rather than a universally available feature.

One customer result worth knowing: the Naked Wardrobe case study reports that Customer Agent resolved 86% of customer queries autonomously and 94% of product-recommendation queries autonomously over a 90-day period. Those are customer-specific results, not a universal benchmark.

What else was announced?

Customer Hub

Customer Hub is live and expanding. Klaviyo is positioning it more clearly as part of the Marketing product line, with additional on-site personalization functionality.

Clicks to Bricks

Clicks to Bricks is a preview focused on location-aware targeting — for example, reaching shoppers who are near a physical location. Public detail is still limited.

Compliance Hub

Compliance Hub is coming soon. The announced direction is a centralized view of compliance health and new requirements, with more detail still to come.

What can you use now?

The status labels are still changing, so here is the practical version:

“Live” does not always mean “available in every account today.” Some of these capabilities are rolling out over the next days or weeks, and public help content may use older status labels while that happens.

How do you connect an AI tool safely?

Start read-only. That is still the right first move.

The Klaviyo MCP server powers the Claude and ChatGPT connections. You can also read the Community's original Claude and ChatGPT overview.

A few settings can narrow what the connection can do:

• read-only=true disables write tools.
• disable-tools-with-user-generated-content=true reduces access to freeform customer text.
• core-tools-only=true reduces the overall tool surface.
• toolsets can narrow the tools exposed through the remote connection by API scope.

Read-only prevents changes, but it does not automatically hide profile data. If your requirement is exact least privilege — for example, read access to Lists and Segments with no Profile access — the more explicit option is a local MCP setup using a private API key with only those scopes.

Klaviyo's current direction favors the remote server for most users, while the local server is a more granular or backward-compatibility path. Either way, review the requested scopes before connecting. The Community has already surfaced how difficult it can be to understand what is being approved, especially for teams without a developer on hand. See How to restrict Claude's data access for the questions customers are asking.

What will Composer cost?

For new packaging, Klaviyo is describing a free tier with 500 credits per month or a $20/month plan with 2,000 credits. Existing paying Composer customers may remain on Early Adoption pricing.

Klaviyo's public pricing page currently describes the free tier as $5 of Composer usage per month, while older help content still references 10,000 free credits for 90 days and $14/month for 2,000 credits. Treat those older numbers as legacy or early-adoption guidance, not universal current pricing.

Actions cost different amounts, and the ranges are estimates that can run higher in practice. The current credit guide estimates:

• Data analysis: 4–25 credits
• Campaign audit: 140–700 credits
• Campaign creation: 100–850 credits

Before you build a workflow around Composer, check Getting started with Composer, what Composer can analyze and create, and How Composer credits work. The documentation is still moving, so avoid hard-coding an assumed cost into a large automation.

What should agencies do differently?

For a remote custom connector, plan on one clearly named connection per client. Klaviyo's MCP Server Guide for Agencies recommends that setup, and connecting several accounts requires a paid Claude plan — Pro, Max, Team, or Enterprise — or ChatGPT developer mode.

That does not mean you need a completely separate workflow for every account. Klaviyo's agency guidance describes connecting multiple client accounts and asking questions across them from one place. The setup is still one connector per client, so account naming, client separation, and approval guardrails matter.

If you manage many accounts, the local MCP server and newer klaviyo-cli beta can also support multiple named account profiles. For scheduled or cross-account automation where you need explicit programmatic control, use the REST API with a separate private key per client.

A few good defaults:

• Name every connector clearly by client.
• Start every connection in read-only mode.
• Keep each client's skills, instructions, and context separate.
• Use a separate private API key per client for API automation.
• Check which AI plan and workspace controls apply to each client's data.

On ChatGPT's Free, Plus, Go, and Pro plans, OpenAI may use information accessed from apps to train its models if the account's “Improve the model for everyone” setting is on. Business, Enterprise, and Edu customers are not used to train models by default. Check this with each client instead of assuming.

What should you watch for?

A few things are still worth treating as open questions:

• Whether global language support is automatically enabled for every existing Customer Agent account.
• Exactly when voice becomes broadly available, and what it will cost.
• Exactly when the new Composer credit packaging and Personalization rollout appear in each account.
• Which Composer edits are available for which accounts and asset types, since public help content still describes a more manual workflow than the K:BOS launch demo.
• Where the current MCP tools stop short. For example, customers are already asking about template access, template relationships, and preserving drag-and-drop editing. See this Community discussion.

None of that should stop you from experimenting with the parts that are live. It does mean you should start with a narrow, read-only use case, document what you approved, and expand only after you understand the results.