Multi-Variant Email Send for AI Agents
A multi-variant email send lets you test more than two subject lines in a single experiment instead of a simple head-to-head. With marginal, your AI agent generates candidates, launches the test, and reads lift metrics — all over the Model Context Protocol.
What a multi-variant email send actually does
Where a classic subject line A/B test compares two options, a multi-variant send fans out across three, four, or more variants so you can explore tone, length, and framing at once. Each recipient slice gets a different subject, and marginal measures which one earns the highest open and click rates within the same test population.
Because variants run side by side under identical conditions, the comparison stays fair. You get within-test lift metrics that tell you how much better the winner performed than the rest, not just a raw rate.
- Generate several subject candidates in one call with generate_variants
- Split traffic across all variants in a single managed send
- Track opens and clicks per variant automatically
- Compare within-test lift instead of guessing from gut feel
Running it through the email MCP server
marginal is a hosted email MCP server, so there's nothing to deploy. Point your client at https://marginal.sh/mcp with a Bearer API key (marg_live_...) and the four tools become available to your agent: generate_variants, launch_test, get_results, and recommend_next.
A typical multi-variant flow chains these tools together. The agent asks for variants, launches the test with launch_test, polls get_results once data accrues, and calls recommend_next to decide what to send to the remaining audience.
- generate_variants — produce multiple subject line options
- launch_test — kick off the multi-variant managed send
- get_results — pull open/click rates and within-test lift
- recommend_next — get a data-backed next-step suggestion
Why agents are good at this
AI agent email marketing shines on multi-variant work because the agent can both create the variants and reason over the numbers. Instead of a human copying subject lines into a dashboard, the agent iterates: read results, refine the angle, and re-test.
This loop fits naturally inside the editors and assistants developers already use, including Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed, and OpenAI Codex. The MCP connection works the same way across all of them.
- Let the agent draft and rank candidate subjects
- Feed results back in to inform the next experiment
- Keep the whole workflow inside your dev environment
Getting started
The free tier covers 100 experiments per month, which is plenty to validate a multi-variant approach before scaling it up. Read the tool reference and example flows in the docs at https://marginal.sh/docs/ to wire your first send.
Start small: a three-variant subject line A/B test on a real list teaches you more about your audience than a dozen untested guesses.
Get started with marginal
marginal is a hosted email marketing MCP server at marginal.sh. Sign up free, create an API key, and connect https://marginal.sh/mcp from Cursor, Codex, or Claude Desktop.
- 100 experiments/month on the free tier
- Four MCP tools: generate_variants, launch_test, get_results, recommend_next
- Listed in the MCP Registry as sh.marginal/mcp