Testing Win-Back Subject Lines with OpenAI Codex and marginal
Win-back campaigns live and die by the subject line. This guide shows how to run a win-back subject line test directly from OpenAI Codex using marginal, the hosted email MCP server, so your AI agent drafts variants, launches a real A/B test, and reports the winner.
Why win-back needs its own subject line approach
Lapsed subscribers have already tuned out your usual cadence, so reactivation hinges on a subject line that feels different from your everyday sends. A win-back subject line test lets you find the angle that actually pulls people back — whether that's a discount hook, a 'we miss you' tone, or a feature update they've never seen.
Running this from OpenAI Codex keeps the whole loop in your editor: you describe the audience, the agent proposes variants, and marginal handles the send and measurement.
- Test re-engagement angles: incentive vs. curiosity vs. social proof
- Compare urgency-driven copy against softer, relationship-led lines
- Measure real open and click lift, not vibes
Generating win-back variants in OpenAI Codex
With marginal connected as an MCP server, OpenAI Codex can call generate_variants to draft a batch of win-back subject lines tuned to your segment. Give it context — how long subscribers have been inactive, what offer you're extending — and it returns distinct options to test rather than near-duplicates.
Because Codex is right in your workflow, you can iterate on the prompt and regenerate until the angles match your brand voice before committing to a launch.
- Ask for 3–5 variants spanning different reactivation hooks
- Refine on tone, length, and emoji usage in plain language
- Hand the best variants straight to launch_test
Launching and reading the subject line A/B test
Once you have your shortlist, launch_test ships the win-back A/B test through marginal's managed sends, then get_results streams back open and click metrics with within-test lift so you can see which subject line is reactivating the most lapsed contacts.
When you want a clear next step, recommend_next tells you which variant to roll out — useful when the agent is driving AI agent email marketing decisions without you eyeballing every number.
- launch_test — start the win-back send and split traffic
- get_results — open/click tracking with within-test lift
- recommend_next — surface the winning subject line to scale
- Free tier covers 100 experiments/month to start
Connecting marginal to OpenAI Codex
marginal is fully hosted, so there's no self-hosted MCP server to maintain — you point OpenAI Codex at https://marginal.sh/mcp and authenticate with a Bearer API key (marg_live_...). From there the four tools are available to your agent.
See the docs at https://marginal.sh/docs/ for setup details and tool schemas, and connect via https://marginal.sh/mcp.
- Endpoint: https://marginal.sh/mcp
- Auth: Bearer API key (marg_live_...)
- Works with Codex plus Cursor, Claude Code, Windsurf, Cline, and more
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