How to Iterate Email Copy With Experiment Data
This how-to guide walks through a tight feedback loop: generate variants, run a subject line A/B test, read the lift metrics, and feed those results back into your next draft. With marginal's email MCP server, your AI agent does each step from one conversation.
Why iterate against experiment data, not opinion
Most copy edits are guesses dressed up as confidence. To actually iterate email copy with experiment data, you need a measured signal after every change — open rates, click rates, and within-test lift between variants. marginal supplies that signal so each revision is grounded in what real recipients did.
Because marginal is a hosted MCP server, your AI agent can run the whole loop without you managing infrastructure. The four tools map cleanly onto the iteration cycle.
- generate_variants — produce candidate subject lines and body copy
- launch_test — start a managed send and subject line A/B test
- get_results — pull open/click tracking and within-test lift
- recommend_next — turn results into a concrete next draft
Step 1: Generate and launch your first round
Start by asking your agent to draft a handful of variants for a single message goal — say, a re-engagement nudge. Keep the variable isolated: if you're testing subject lines, hold the body steady so the lift you measure is attributable.
Once you have variants you like, launch the test. marginal handles the managed send and splits delivery across the variants automatically.
- Prompt: "Generate 4 subject line variants for a 90-day re-engagement email."
- Then: "Launch a subject line A/B test with these against my dormant segment."
- Confirm the test ID the agent returns so you can fetch results later
Step 2: Read results and feed them back
After the send accumulates engagement, call get_results to see opens, clicks, and the within-test lift of each variant relative to the others. This is the data that drives the next edit — a winning subject line tells you which angle, length, or tone resonated.
Hand those findings to recommend_next. The tool reads the result set and proposes the next iteration: sharper variants that double down on the winning pattern, or fresh angles when the field was flat. That closes the loop, and you repeat from Step 1.
- Ask: "Show me results and lift for test <id>."
- Then: "Based on those results, recommend my next round of variants."
- Iterate until lift plateaus, then change the variable you're testing
Running the loop from your editor
This marginal MCP tutorial works in any MCP client — Cursor, Claude Code, Claude Desktop, Windsurf, Cline, Continue, Zed, or OpenAI Codex. Point the client at https://marginal.sh/mcp with your Bearer API key and the four tools become available to your AI agent for email marketing.
The free tier covers 100 experiments per month, which is plenty for running several iteration cycles. See https://marginal.sh/docs/ for tool schemas and example prompts.
- Endpoint: https://marginal.sh/mcp
- Auth: Bearer key in the form marg_live_...
- Free tier: 100 experiments/month
- Docs: https://marginal.sh/docs/
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