Understanding marginal launch_test parameters
The launch_test tool is how an AI agent actually kicks off a subject line A/B test on marginal. This guide breaks down the marginal launch_test parameters so your agent can configure sends correctly the first time.
Where launch_test fits in the workflow
marginal exposes four MCP tools to AI agents: generate_variants, launch_test, get_results, and recommend_next. The launch_test call is the moment a planned experiment becomes a live send. Before reaching for it, most agents generate candidate subject lines with generate_variants, then pass the strongest ones into launch_test.
Because marginal is a hosted email MCP server, you don't manage send infrastructure or worry about a self-hosted process. The agent supplies parameters, marginal handles the managed send and the open/click tracking behind the scenes.
- generate_variants — produce subject line candidates
- launch_test — start the subject line A/B test
- get_results — read open/click and within-test lift metrics
- recommend_next — get a suggested follow-up experiment
Core launch_test parameters
When an AI agent calls launch_test, it describes the experiment: the variants under test, the audience, and the body content that ships with each subject line. The exact parameter names and shapes are documented at https://marginal.sh/docs/, but conceptually these are the inputs you'll always reason about.
- Subject line variants — the competing lines being measured against each other
- Email body — the message content sent with each variant
- Audience — the recipient set the test runs against
- Test identification — a name or label so get_results can find the run later
Tips for AI agent email marketing
In AI agent email marketing flows, the value of launch_test comes from keeping variants comparable. Hold the body constant and vary only the subject line so the lift you measure reflects the subject, not other changes. Two to four variants per subject line A/B test is usually enough to get a clear signal without splitting your audience too thin.
After a test runs, poll get_results for within-test lift metrics rather than guessing. Let recommend_next propose the follow-up so the agent iterates on real open and click data instead of assumptions.
- Vary the subject line only; keep the body fixed
- Use a manageable number of variants to preserve statistical signal
- Label tests clearly so results are easy to retrieve
- Stay within the free tier of 100 experiments/month while prototyping
Connecting and authenticating
launch_test is available once your MCP client is pointed at the marginal endpoint at https://marginal.sh/mcp and authenticated with a Bearer API key (marg_live_...). Supported clients include Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed, and OpenAI Codex.
With the connection in place, any of those agents can call launch_test with the parameters above and immediately manage subject line experiments end to end.
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