Integration · Eppo
Eppo is the warehouse-native experimentation platform used by Twitch, DraftKings, and Perplexity.
Strong product. No first-party MCP server as of May 2026. Programmatic access goes through the Eppo REST API. This page is included for completeness so the experimentation category covers the platforms PMs actually evaluate, not just the ones with shipping MCPs.
What Eppo does well
Eppo runs experiment math on the team's data warehouse (Snowflake, BigQuery, Redshift, Databricks). Event data never leaves the stack to power the statistics. For data-team-led organizations that already own the warehouse, that architecture means experimentation joins the rest of the data work rather than living in a vendor silo.
The product covers contextual bandits for adaptive allocation, lifecycle experimentation for retention work, and explainability features that let data scientists trace any reported lift back to the underlying SQL. The customer list (Twitch, DraftKings, Perplexity) skews toward data-mature product orgs, which is consistent with the architecture.
Why no MCP yet
Eppo's primary audience is data scientists and experimentation analysts. Those teams typically work in notebooks, dashboards, and the warehouse itself; the push toward agentic workflows has come from product managers and engineers, who are not Eppo's buyer. That likely explains why an MCP has not been a Q1 or Q2 2026 priority.
The signal is that the demand is growing. Several customers have asked publicly for an Eppo MCP, and the category as a whole has shifted in this direction over the last six months. A first-party Eppo MCP is plausible in the next quarter or two; this page will be updated when one ships.
How to integrate today
The supported path is the Eppo REST API. Standard auth via API key. The endpoints cover experiment definitions, assignment configs, results, and metric definitions, so the agent can read and write the platform from outside an MCP runtime.
Teams that want agent-driven Eppo work today can build a small MCP adapter on top of the REST API. The Anthropic MCP SDK plus 200 lines of tool wrappers covers the most common operations (list experiments, get results, create experiment, update assignment). It is not a small lift, and it locks the team into maintaining the adapter, but it is a viable path while the first-party MCP catches up.
For most teams the right call is to use the REST API directly from CI and scripts and to wait for the first-party MCP for agent workflows.
Visual demonstration
Primary metric: d7_retention (warehouse SQL)
| Variant | Metric | Lift | 95% CI | p |
|---|---|---|---|---|
| Controlcontrol | 41.20% | baseline | — | — |
| Treatment | 43.05% | +4.49% | [1.40, 7.60] | 0.005 |
Result computed on Snowflake. Lift held in mobile and desktop subgroups. Recommendation: ship; lifecycle effect worth tracking past launch.
CLI alternative
No first-party CLI, no first-party MCP. The REST API covers experiment definitions, assignment configs, results, and metric definitions. Standard API key auth. For most teams this is the supported path until the first-party MCP ships.
Pairs with these skills
The forthcoming experiment-design skill applies cleanly to Eppo because the framework is platform-agnostic. The implementation pattern differs: the agent calls the Eppo REST API rather than an MCP. Skill landing pages and SKILL.md sources land in subsequent dispatches.
This page will be updated when Eppo ships a first-party MCP. Useful for visitors who land here looking for one. The category-level signal is that demand is growing across customers; a near-term release is plausible.