Klaviyo’s AI marketing agent has a new working pattern. You no longer have to open it and ask. You can give Composer a standing job, like “every Monday, review our top flows and flag anything slipping,” and it runs in the cloud whether or not anyone on your team is logged in.
That is the useful part of this release, and it is genuinely useful. Most lifecycle teams have a weekly audit that slips the moment a launch lands on the calendar. A tireless junior analyst that checks flows on Monday morning is worth having.
It also changes the questions you need answered before you switch it on. Who decided what the night shift checks? Who sees the bill, given that Klaviyo does not show the cost of a job until after it runs? And who else can reach the same account while Composer politely waits for your approval?
A note on evidence. This is a primary-source review. I have not used Composer in a Klaviyo account for this piece. Everything below comes from Klaviyo’s October 13 launch post, the Klaviyo Community release recap, seven Klaviyo Help Center articles (most updated October 14), the Composer product and pricing pages, the September K:BOS announcement, and Klaviyo’s developer documentation for its MCP server, all read on October 15. I cross-checked against an agency account published on Klaviyo’s blog and an independent BizStack write-up.
What it is
Composer is the agent built into Klaviyo, opened from a sparkle icon anywhere in the app. Klaviyo’s help center splits what it does into two modes.
The first is analyze and audit. You ask a question about your account in plain language, or ask it to review something you already built: a flow, a campaign before send, a single email or text message, a segment, a signup form. Audits come back as a prioritized list of findings. They are advisory, and Klaviyo is explicit that “an audit does not change your work.”
The second is create. Composer drafts email, SMS, MMS and push campaigns from a goal, drafts new flows with their messages, and drafts segments with a profile count for you to check. Drafts open in a preview, then move into the normal Klaviyo editors for final edits.
Across both modes the help center repeats one rule: “Composer does not send or publish campaigns, flows, or messages without your sign-off.” Scheduled tasks are covered by that too. Klaviyo says recurring analysis is “separate from scheduling marketing messages to your customers.”
What changed this month
Klaviyo dates the release two ways. The October 13 blog post by Elias Torres says “starting today, Composer is now available to every Klaviyo customer.” The Community recap posted the same day says Composer has been out of preview and available to every account “as of October 6.” Either way, it is no longer a preview, and every account gets a free plan with nothing to switch on.
Three capabilities are new or newly broad.
Cited analysis on a real query engine. Before this release, according to the Community recap, Composer could only answer questions its pre-built skills covered. Ask something outside that list and it said it could not help. Now Klaviyo translates a plain-language question into SQL, runs it against your account’s data using Klaviyo’s own definitions of terms like revenue, and hands back the answer with the definition and the query. Results can be compared with benchmarks Klaviyo collects across more than 205,000 customers.
Scheduled tasks. Recurring jobs on a cadence you set, running in the cloud while you are away. Klaviyo’s examples are a Monday flow review and a first-of-the-month comparison of lifecycle performance across markets.
Skills. Your team’s process, written once and reused. At K:BOS in September, Klaviyo described custom Composer Skills as saved approaches like subject line standards, brand guidelines or content hierarchy, so you do not repeat them in every prompt.

What works
The approval gate is consistent. Every Composer help article that describes what it does repeats the same rule in slightly different words: Composer drafts, audits and proposes, and a person approves anything customer-facing. Even the “Apply recommendations” option on an email template audit prepares changes in a preview and waits for you to click Update. That is a clear, checkable promise.
Audits are the strongest use. Klaviyo published an account from Sticky Digital, a retention agency, on its own blog, so read it as a practitioner story rather than an independent test. The agency says it uses Composer more for analysis and QA than for creation, with two QA passes on every client: one by Composer, one by a person. Its example is concrete. A welcome series reached 165 people with a 38.9% open rate, zero clicks and zero revenue. Composer flagged 17 issues, led by broken click tracking, subscribers stalling after the first email, a subject line typo on three messages including both A/B variants, a 13.7% bounce rate on one email, and two flows checking the same profile property stored as text in one and as a number in the other. That is exactly the kind of quiet breakage a weekly audit should catch.
Answers show their work. The SQL engine returns the definition and the query behind a number. That matters because two chats that guess differently at “revenue” give you two different answers. The Community post still advises spot-checking early answers against a number you already trust, which is good advice from the vendor’s own manager.
Image controls exist. Under Settings, then Composer, anyone with permission to edit account settings can separately turn off AI image generation and AI editing of existing images for the whole account. Brands with strict photo rules can close that door on day one.
What breaks or is limited
You learn the price after the job. The plan article is blunt: “No exact cost is shown before an action runs.” Klaviyo publishes estimated ranges per action and warns that actual usage can go above the upper limits. Analyze data runs about $0.35 to $0.90. Auditing a flow, form or template runs $0.10 to $4.50. Creating a campaign runs $0.10 to $10.00. Revisions use capacity like the first draft.
Put that next to a schedule. Klaviyo does not publish a separate price for scheduled tasks, but a scheduled audit is still an audit, so I would assume it draws from the same table. Here is an illustration using only Klaviyo’s published upper bounds, not a measured bill: a Monday job that audits five flows, each landing at $4.50, costs about $22 a week. Over a month that is several times the $20 paid plan. Real jobs may cost far less. The point is that nobody on your team sees the number until the job has already run, and the job runs while they are asleep.
Overages buy in $10 blocks. On the paid plan, flexible overages automatically purchase 1,000 more credits for $10 when you hit the limit. Klaviyo charges the full $10 even if you use part of it, and several blocks can land in one cycle. You can set the preference to None so credit actions stop at the limit instead. Only someone with billing permission can change it, and unused capacity does not roll over.
The gate belongs to Composer, not to your account. This is the limitation that matters most. Klaviyo’s own K:BOS post says its headless MCP layer lets an agent “read, write, and go live the same way you would inside Klaviyo, without anyone opening Klaviyo at all.” The developer docs list over 260 MCP tools, and the remote server includes send_campaign, described as “send a campaign to its configured audience,” along with update_flow for flow status, bulk_unsuppress_profiles and marketing subscribe and unsubscribe tools. The Claude connector’s OAuth grant includes read and write on campaigns, flows, profiles, segments and subscriptions. Klaviyo’s FAQ says it does not currently restrict which AI clients can connect, and that the Klaviyo UI does not show which user authorized the MCP app. API Analytics does log which tools were called.
None of that makes Composer unsafe. It means “Composer never sends without approval” is a sentence about one window. If your team also connects Claude, ChatGPT or an in-house agent to the same account, that sentence does not cover them.
Definitions are Klaviyo’s until you change them. The SQL engine uses Klaviyo’s definitions. Sticky Digital notes that it counts on-site activity as “active,” while Klaviyo’s default counts only email or text engagement. If your house definition of active, revenue or lapsed differs, a scheduled report will faithfully repeat the wrong one every Monday.
The docs do not agree on everything. The Composer product page lists email, SMS, push, WhatsApp and RCS, with availability varying by account. The help center’s create articles document email, push, SMS and MMS today. Composer is also weaker on image-heavy or fully image-based templates, by Klaviyo’s own note, and Klaviyo says plainly that its output “does not guarantee accuracy.” On Skills, I found no help article explaining how Composer Skills are permissioned or carried across separate client accounts. The help article titled “How to create a Skill” covers Customer Agent skills, which is a different product.

Who it is for right now
It fits a solo retention marketer or small team running one Klaviyo account with a backlog of audits nobody gets to. It fits agencies that already run a written QA checklist and want a first pass before the human pass, which is how Sticky Digital describes its use. It fits brands with multiple markets that want the same monthly comparison every month without someone rebuilding the export.
It fits badly as an unattended creator. Klaviyo itself tells you never to send a draft as-is. It fits badly for teams that have not agreed on core definitions, because a cited answer to a badly defined question is still the wrong answer, delivered with confidence. And it fits badly in accounts where several people and outside agents already hold write access with no record of who connected what.
Pricing and limits
Every account gets a free Composer plan: 500 credits per cycle, which the pricing page expresses as $5 of Composer usage a month. It resets each cycle and is not a trial. At 100%, credit actions stop until the reset or an upgrade. Support and general Klaviyo questions are free, and Klaviyo’s legacy credit article says actions that fail or time out on Composer’s side do not use credits.
The paid plan is $20 a month for 2,000 credits. Klaviyo’s two price tables line up at one cent per credit, so the $10 top of the create-a-campaign range is 1,000 credits: twice the free plan’s whole cycle and half the paid plan. Overages are $10 per 1,000 credits, bought automatically only if you turn flexible overages on. Usage notices appear at 80%, 90% and 100%. Early adopters on legacy terms had a 10,000-credit, 90-day trial and a $14 entry plan, which Klaviyo labels legacy pricing.
What to do this week
Composer is worth turning on. Turn it on like you would onboard a new analyst who works nights and bills by the job.
- Write a scheduled-job list of no more than three recurring tasks, each with an owner, a cadence and a named reader for the output. Start with audits, not creation.
- Decide on the meter before the first schedule. Leave flexible overages at None for the first cycle, check the usage panel after each scheduled run for two weeks, and only then decide whether the paid plan or overages make sense.
- Put your house definitions in writing: active, lapsed, revenue (gross or net, refunds in or out), and your maximum safe segment size. Give them to Composer as context or a Skill, and check its first SQL answers against a number you already trust.
- Inventory every outside agent connected to the account through MCP. Use API Analytics to see what each one calls. Set any connection that only needs reporting to read-only, and write down which humans may connect a write-capable client.
- Agencies: confirm in writing who pays for Composer usage on each client account, who holds billing permission, and whether your Skills live in your account or the client’s.
- Decide your AI disclosure line and placement now, using Klaviyo’s examples as a starting point and your counsel as the final word, and turn off AI image generation if your brand does not use it.
Composer’s promise is that it will always ask before it talks to your customers. Take it at its word. Then make sure the night shift, the meter and every other agent in the account are asking too.




