Feature Release Email Subject Test
A feature launch only matters if people open the email. This guide shows how to run a feature release email subject test with marginal — the hosted email MCP server that lets your AI agent generate subject variants, launch a real A/B test, and pick the winner based on opens and clicks.
Why subject lines decide your launch
Feature announcements live or die in the inbox. Two phrasings of the same release can produce very different open rates — one teases the benefit, another names the feature outright, and only data tells you which lands with your list.
Instead of guessing, a feature release email subject test puts a few candidate lines in front of a sample of your audience and measures real engagement before you commit to the full send.
- Benefit-led: "Stop copy-pasting reports — here's the fix"
- Feature-named: "Introducing scheduled exports"
- Curiosity: "The thing you asked for is live"
- Each gets measured on open and click-through, not vibes
How marginal runs the test
marginal is an email MCP server your AI agent connects to over https://marginal.sh/mcp. Rather than wiring up dashboards by hand, you describe the release and let the agent drive the workflow through four MCP tools.
The flow maps cleanly onto a feature launch: draft candidates, split-test them, read the lift, and act on the recommendation.
- generate_variants — produce subject line options from your release notes
- launch_test — start a subject line A/B test on a sample audience
- get_results — pull open/click tracking and within-test lift metrics
- recommend_next — get the winning subject to send to the rest of the list
Driving it from your AI agent
Because marginal speaks MCP, AI agent email marketing happens inside the tools you already use. Connect from Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed, or OpenAI Codex with a Bearer API key (marg_live_...), and your agent can run the entire subject test conversationally.
A typical prompt: "We're shipping scheduled exports next Tuesday — generate four subject variants, launch a test to 20% of the list, then recommend the winner." The agent calls the MCP tools in sequence and reports back the lift.
- Hosted — no self-hosted server to run or maintain
- Free tier covers 100 experiments per month
- Managed sends with built-in open and click tracking
- Docs and tool references at https://marginal.sh/docs/
Putting it into practice
Treat each launch as a repeatable experiment. Keep a small holdout for the test split, give every variant the same body so you isolate the subject's effect, and let recommend_next decide the broadcast line once results stabilize.
Over several releases you build an intuition — backed by lift data — for which subject styles your audience opens, turning every feature announcement into a slightly better one.
- Test 3–4 distinct angles, not minor word swaps
- Hold the email body constant across variants
- Send the winner to the remaining audience automatically
- Log results to compare patterns across launches
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