Eppo MCP server: it is Datadog Experiments now

Why searching for an Eppo connector finds nothing, which 22 Datadog tools replace it, and what an agent can start, conclude or cancel with them.

Last verified 1 September 2026 · from geteppo.com, its documentation sitemap and Datadog's tool reference

This page is one of 90 in a directory of monitoring and developer tools. Each page checks what an AI coding agent can get out of the tool through its MCP server, the connector that lets the agent query the tool directly.

What is the Eppo MCP server?

Eppo was an experimentation platform: define a metric, split traffic with a feature flag, and get a statistically defensible answer about which variant won.

Datadog acquired it, and the product now ships as Datadog Experiments alongside Datadog Feature Flags. The concepts survive: experiments, variants, primary metrics, subject types (what each unit of traffic is, such as a user or a session) and allocations (how that traffic is split).

Reaching them from an agent means connecting Datadog's hosted MCP server. There is no separate Eppo endpoint, and there never was.

What a search for an Eppo connector actually finds:

Eppo's site still resolves independently and still carries its own product pages. The acquisition banner is what explains the missing connector.

Experimentation sits in release, and the surrounding Datadog server reaches much further.

For an agent, that means the whole experiment lifecycle is reachable, from creating one to declaring a winner. The statistical judgment stays in Datadog; what the agent gets is a verdict field and a health summary.

What can an AI agent do with Eppo?

What connecting costs before the first question

A context window is the amount of text a model can hold at once. Eppo publishes no server of its own, so what is measured here is Datadog Experiments, the route this page recommends instead. Twenty-one tools cost 1,786 to 2,082 tokens on Claude.

The figure belongs on this page because the substitution is the finding. An agent reaching Eppo's experiment data pays this to do it through someone else's connector.

Encodings: GPT-5.6, GPT-5.5 o200k_base; GPT-4, GPT-3.5 cl100k_base; Claude Haiku 4.5 to Opus 5 count_tokens. Counted from the 22 Experiments and Feature Flags entries in Datadog's tool reference. That reference is the only public inventory.

The hosted server needs real keys. So no live tools/list was captured, and no input schemas are available. These figures cover each tool's name and description only. They are a floor, not the figure.

The Claude columns are measured through the API's count_tokens endpoint with an empty schema per tool. So they cover the same content as the GPT columns. The range there is two tokenizer generations. Haiku 4.5 and Opus 4.6 are at the low end. Opus 5 and Sonnet 5 are at the high end. The gap between them is about a fifth.

Eppo MCP server limits: where answers come back incomplete

The first item below sends readers looking for a connector that does not exist. Read it before the caveats about live experiments.

How to configure Eppo MCP for agents

1. 1

   Stop looking for an Eppo endpoint

   There is none, and the documentation confirms it: 238 URLs with no MCP page.

2. 2

   Connect Datadog instead

   The hosted server at mcp.datadoghq.com/v1/mcp is the whole route.

3. 3

   Narrow the toolsets

   Twenty-two of the 265 are relevant here, and a full connection spends context on the other 243.

4. 4

   Read the diagnostics before the results

   get_experiment_diagnostics returns a sample ratio mismatch check. A sample ratio mismatch means the traffic each variant actually received does not match the split you configured, which is the standard reason a result should not be believed.

5. 5

   Decide who may conclude an experiment

   conclude_experiment records a permanent winning variant decision.

Paste this into your agent’s instructions

An experiment result is a statistical claim, and the tools return a verdict field that reads like a plain answer. The block keeps the health check attached to it.

When you answer using Datadog Experiments MCP tools,
state:
- The experiment name and the primary metric the verdict
  refers to.
- Whether get_experiment_diagnostics was called, and what it
  said about sample ratio mismatch. Do not report a verdict
  without it.
- Whether the call changed anything. create, start, cancel
  and conclude all change a live experiment.
Never state that data does not exist. State which tool you
called and what it returned.
Write your answer in ASD-STE100 Simplified Technical
English. Use short sentences with one idea in each.

Do you need the MCP server at all?

Not always, and never under the Eppo name. The Datadog UI remains the direct route for reading a result, and it shows the confidence intervals a prose answer flattens.

For experimentation without the rest of Datadog, Statsig publishes its own server and covers gates, experiments and autotunes directly.

For flags alone, Flagsmith and Harness FME are narrower connections than a 265-tool server.

Eppo's own API documentation is still published, and remains the route for anything scheduled against the platform.

With MCP, a result, its health check and the flag behind it arrive in one conversation, which is the sequence a careful reading of an experiment needs anyway.

Connect Datadog, filter to Experiments and Feature Flags, and gate the eight write tools.

Eppo MCP server setup

Nothing installs under the Eppo name. Eppo is a Datadog company, and its coverage arrives through Datadog's hosted MCP server rather than through a server of its own.

That has a practical consequence worth knowing before you start: an Eppo login on its own reaches no MCP server at all. You need a Datadog account.

There is nothing to download. The server is hosted, and the endpoint answered HTTP 401 to an unauthenticated call, which is what a live gated endpoint should do.

claude mcp add --transport http datadog \
  https://mcp.datadoghq.com/v1/mcp

{"mcpServers":{"datadog":{
  "url":"https://mcp.datadoghq.com/v1/mcp",
  "headers":{"DD_API_KEY":"<your-api-key>",
             "DD_APPLICATION_KEY":"<your-app-key>"}}}}

{"mcpServers":{"datadog":{
  "url":"https://mcp.datadoghq.com/v1/mcp",
  "headers":{"DD_API_KEY":"<your-api-key>",
             "DD_APPLICATION_KEY":"<your-app-key>"}}}}

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

curl -s -o /dev/null -w "%{http_code}" -X POST \
  https://mcp.datadoghq.com/v1/mcp        # 401

Read-only access, permissions and security

• Eight of the 22 tools write

  They create experiments and flags, start and cancel runs, conclude a winner, and change allocations in a named environment.

• conclude_experiment is permanent

  Datadog's own description calls it a permanent winning variant decision, which is not a state an agent should reach unattended.

• Allocation changes move real traffic

  update_datadog_feature_flag_environment and sync_datadog_feature_flag_allocations act on a specific environment, production included.

• Cancelling requires a reason, and the model writes it

  cancel_experiment takes a required reason, so the audit trail records whatever the agent typed.

• The credential is a Datadog credential

  It carries whatever the Datadog account can reach, which is far more than experiments.

Troubleshooting

You cannot find an Eppo MCP server

There is none. Eppo's documentation sitemap held 238 URLs on 25 August 2026 with no MCP page, and the Eppo-exp GitHub organization has no MCP repositories.

An Eppo login does not authenticate the server

The server is Datadog's. It authenticates with Datadog credentials, and the Eppo product now ships as Datadog Experiments.

The experiment will not start

start_experiment requires a linked flag with an active allocation, a subject type and a primary metric. link_feature_flag_to_experiment is the tool that supplies the first of those.

The results look wrong

Call get_experiment_diagnostics first. It returns a health summary including sample ratio mismatch status, which is the usual explanation for a result that does not make sense.

The analysis pipeline failed

diagnose_experiment_run_failure exists for exactly this. It diagnoses why the latest or a specific analysis run failed and returns the root cause.

Too many unrelated tools appear

The server held 265 tools on 1 September 2026. Only the Experiments and Feature Flags toolsets are relevant here, so filter to them.

Eppo MCP server: Reference

What the vendor documents

There is nothing to report about an Eppo connector, because none exists. What follows is the evidence for that, and the shape of the coverage that replaced it.

“Eppo is now Datadog Experiments.”

“Concludes a running experiment with a permanent winning variant decision.”

“Returns a health summary for an experiment before interpreting results: sample ratio mismatch (SRM) status.”

Eppo's documentation sitemap was fetched on 25 August 2026 and its 238 URLs were searched for MCP, agent and assistant, with no match. The Eppo-exp GitHub organization was searched for MCP repositories on the same date and returned zero. Tool names and descriptions come from Datadog's own tool reference, read 18 August 2026.

Should you connect an agent to Eppo?

Yes. Best for reading a verdict together with its health check, with every write tool gated.

Strongest fit

Teams already on Datadog who ran experiments on Eppo and want an agent reading results and flags.

Main advantage

The full experiment lifecycle is reachable, and the diagnostics that decide whether a result is trustworthy are a tool rather than a footnote.

Main weakness

Nothing carries the Eppo name, so the connector is 22 tools inside a 265-tool server built for something wider.

Operational risk

conclude_experiment writes a permanent winning variant decision, and allocation tools move traffic in a named environment.

Frequently asked questions