MCP Server
Backstroke MCP Server
The predictive layer for your marketing AI.
Connect Backstroke to Claude, or any AI assistant, and get subject lines, heroes, and content predicted to perform. Learned from 10,000 ecommerce brands, returned right in your prompt.
Overview
The Backstroke MCP Server connects your AI assistant to Backstroke's prediction models. Most marketing connectors give an assistant read access to your own account data. Backstroke returns something your account cannot produce on its own: a read on what will actually perform, learned from real email outcomes across 10,000 brands.
You ask in plain language. Backstroke answers right in your prompt, and with your approval it can write the winner back into your email platform. Your work stays in the assistant you already use, and Backstroke stays the prediction engine behind it.
How it works
Backstroke MCP is a hosted endpoint your assistant connects to. You sign in once. From then on, your assistant can call Backstroke's tools whenever you ask.
- You ask. Describe what you are working on, such as subject lines for an email or which hero wins for your audience, in natural language.
- Backstroke predicts. Your request runs against models learned from outcomes across 10,000 brands, not your account alone.
- You decide. The answer comes back in your prompt. Anything that writes to your email platform pauses for your explicit approval first.
You connect once. As new tools ship, they appear in your assistant automatically, with no reinstall and no reconnect.
The tools
Backstroke's tools arrive in your assistant in two groups, split by what they are allowed to touch. Read-only tools are always allowed. Write tools ask for your approval every time. Tools marked Live work today. Tools marked Coming are on the way.
Read-only
Always allowedThese tools read and predict. They never change anything in your systems.
Generate a set of subject lines and preheaders predicted to perform, each with a one-line why.
The same set, tuned per segment and written in your brand voice.
Read an existing campaign's real content, its hero, products, and copy, so every suggestion is about that email and not just its name.
Drop in candidate hero images and get them ranked by predicted performance, with the audience each one wins with.
Score heroes against your actual audience and get the winner, with the segments it wins with.
Generate the hero or heroes predicted to perform, either a single winner or a set routed to different cohorts as one personalized block.
Read your results and return campaign lift with a board-ready roll-up.
Write
Approval requiredThese tools change something in your email platform. Your assistant asks before each one, every time. Nothing sends on its own, and writes create drafts you review.
Write your per-segment subject lines into Klaviyo as draft campaign variants, after you confirm.
Turn a Backstroke campaign predictive by attaching the winning hero set and updating the email.
Build a Klaviyo content block that routes the predictive heroes per segment.
Teach the agent a rule once, such as never say discount, and it updates your brand config so every future suggestion respects it.
Prediction, not generation
Backstroke works in two verbs, and the distinction matters. Subject lines are generated. You get a set predicted to perform, written from patterns across 10,000 brands, rather than a ranking of lines you already have. The edge is the model underneath, not a guaranteed winner.
Heroes are scored. Candidate images are ranked against a model of your actual audience, and you get the winner along with the segments it wins with. That is the thing a generic generator cannot do, because it has no read on your list. Hero scoring returns a ranked read against your audience, with expected lift added once it is validated against real results.
What you can ask
With the Live tools, you can ask your assistant things like these today:
Subject lines
Heroes and content
Activate
Generating predictive heroes and pushing the winner live are coming next, and will answer the same way, right in your prompt.
Staying in control
The agent proposes. You commit. The permission model is built into how the connector works:
- Read-only by default. Every prediction tool only reads and answers. Nothing in your systems changes.
- Writes always ask first. Anything that touches your email platform pauses for your explicit approval, every time. Writes create drafts, and nothing sends on its own.
- Nothing stored. Your campaign content is processed to answer your question, not retained.
- Connect once, revoke anytime. Sign in with OAuth. Every call resolves to your account, and you can disconnect whenever you want.
- SOC 2 Type II compliant. Backstroke maintains SOC 2 Type II controls across its platform.
Roadmap
The Live tools are the first rung. What is ahead:
- Any assistant. Claude today, other AI assistants next, from the same connection.
- The full hero workflow. Read the campaign, score heroes for your audience, and generate the predictive set, then activate the winner where you work.
- The whole play in one command. Ask once and the agent runs the sequence for you, reading, scoring, generating, and writing, holding for your approval at each write.
Access
Customers get first access.
Backstroke MCP access is rolling out to Backstroke customers first. Fill out the form and we will walk you through getting connected.
Request access →