Working with marginal experiment limits through the API
If you're building automated subject line tests with an AI agent, you'll want to know how experiment counting works before you scale. This guide explains the marginal experiment limits API behavior, what counts against your quota, and how to plan your sends.
What counts as an experiment
In marginal, an experiment is a single subject line A/B test launched against an audience. Each call to launch_test starts one experiment that consumes one slot from your monthly allowance. Generating variants or pulling results does not count toward the limit on its own.
The free tier includes 100 experiments per month. That's enough to run several subject line A/B tests per day while you're prototyping an AI agent email marketing flow.
- launch_test → consumes one experiment from your monthly quota
- generate_variants → drafts subject lines, no quota cost
- get_results → reads open/click and lift metrics, no quota cost
- recommend_next → suggests a follow-up test, no quota cost
Checking limits from the API
Because marginal is a hosted email MCP server, your agent talks to it over the endpoint at https://marginal.sh/mcp using a Bearer API key (marg_live_...). The MCP tools return structured results, so your agent can read what happened on each launch_test call and adjust behavior as it approaches the monthly cap.
A practical pattern is to have the agent batch related subject line tests, then poll get_results before launching new ones. This keeps your experiment count tied to tests that actually delivered learnings rather than redundant launches.
- Endpoint: https://marginal.sh/mcp
- Auth: Bearer API key beginning with marg_live_
- Registry identifier: sh.marginal/mcp
- Full reference: https://marginal.sh/docs/
Staying within the limit at scale
The quickest way to burn through experiments is to launch a new subject line A/B test for every minor copy tweak. Instead, let your agent consolidate hypotheses into fewer, sharper tests and use recommend_next to decide whether another experiment is worth the slot.
Tracking within-test lift helps here: once one variant clearly wins, there's no need to spend another experiment re-running the same comparison.
- Group similar variants into a single launch_test call
- Read lift metrics before deciding to launch again
- Use recommend_next to avoid low-value follow-up tests
- Schedule tests across the month rather than front-loading them
Where this fits in an AI agent workflow
marginal connects to MCP clients including Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed, and OpenAI Codex. Your agent can generate variants, launch a managed send, and report back on opens and clicks — all while respecting the experiment limits exposed through the API.
For teams running AI agent email marketing, the 100-experiment free tier is a sensible budget for validating your flow before committing to higher volume.
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