A Practical Email Experimentation Framework
An email experimentation framework gives your team a repeatable loop for proposing, testing, and learning from email variants. marginal turns that loop into MCP tools your AI agent can call directly, so subject line A/B tests run end-to-end without leaving your editor.
What an experimentation framework needs
Most teams treat email tests as one-off events: someone writes two subject lines, eyeballs the open rate a day later, and moves on. A real framework formalizes each stage so results compound over time and decisions stay grounded in measured lift rather than gut feel.
- Hypothesis: a clear claim about what change should move opens or clicks
- Variant generation: enough distinct options to make a test worthwhile
- Controlled send: a managed split so traffic is fairly distributed
- Measurement: open/click tracking with within-test lift between arms
- Iteration: a recommendation for what to try next based on results
How marginal implements the loop
marginal is a hosted email MCP server that maps each stage of the framework onto a tool your AI agent can invoke. Because it runs as a managed endpoint, there is nothing to self-host — you point a client at https://marginal.sh/mcp and authenticate with a Bearer API key.
The four tools form a closed loop. An agent drafts options with generate_variants, ships a subject line A/B test through launch_test, reads open and click metrics from get_results, and asks recommend_next for the next experiment to queue.
- generate_variants — produce candidate subject lines from a brief
- launch_test — start a managed, tracked split send
- get_results — pull open/click tracking and within-test lift
- recommend_next — suggest the next hypothesis to test
Driving the framework from an AI agent
AI agent email marketing works best when the agent can both reason about copy and act on real send data. With marginal connected over MCP, an agent in Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed, or OpenAI Codex can run the full cycle conversationally — describe a campaign, get variants, launch a test, and review lift in the same thread.
This keeps experimentation continuous instead of episodic. The agent already has context on past results, so each new subject line A/B test builds on what worked before rather than starting from scratch.
Getting started
The free tier includes 100 experiments per month, which is plenty to validate the framework on a real list before scaling. Add the email MCP server to your client, run a first generate_variants call, and launch a test to see open/click tracking and lift flow back automatically.
See the docs at https://marginal.sh/docs/ for tool schemas and client setup, or connect directly to the endpoint at https://marginal.sh/mcp.
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