Skip to main content
Glama
Sign In
unmeshmali25

experiment-mcp

by unmeshmali25
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Hybrid

Server Configuration

Describes the environment variables required to run the server.

NameRequiredDescriptionDefault
EXPERIMENT_MCP_JWKS_URLNoURL of the JWKS used to validate bearer JWT signatures.
EXPERIMENT_MCP_JWKS_PATHNoPath to a local JWKS file used to validate bearer JWT signatures. Offline alternative to EXPERIMENT_MCP_JWKS_URL.
EXPERIMENT_MCP_ISSUER_URLNoIssuer URL for bearer JWT validation. Tokens must have a matching issuer. Must use HTTPS unless the host is localhost/127.0.0.1.
EXPERIMENT_MCP_RESOURCE_URLNoResource/audience URL for bearer JWT validation. Tokens must have aud/resource exactly equal to this value. Must use HTTPS unless the host is localhost/127.0.0.1.
EXPERIMENT_MCP_ALLOWED_HOSTSNoComma-separated list of extra Host values allowed by DNS rebinding protection.
EXPERIMENT_MCP_ALLOWED_ORIGINSNoComma-separated list of extra Origin values allowed by DNS rebinding protection.
EXPERIMENT_MCP_ALLOW_INSECURE_NO_AUTHNoSet to 1 to start HTTP transport without token validation. Cannot be combined with issuer or JWKS settings. Intended only for local use; do not expose the process.

Instructions

Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.

This server publishes no instructions, or was last inspected before Glama recorded them.

Capabilities

Features and capabilities supported by this server

Protocol revision2025-11-25

CapabilityDetails
tools
{
  "listChanged": false
}
prompts
{
  "listChanged": false
}
resources
{
  "subscribe": false,
  "listChanged": false
}
completions
{}

Tools

Functions exposed to the LLM to take actions

NameDescription
power_calcA

Estimate equal-allocation sample size or MDE for a two-sided test.

    Provide exactly one of mde or n_per_arm. The normal approximation does
    not include peeking, multiple metrics, or a finite-population correction.
assign_cellsA

Assign units to arms with a salted SHA-256 bucket.

Arms are ordered by name. The same salt and unit id always map to the same arm. This does not write an experiment record.

analyze_liftA

Estimate absolute and relative lift with a confidence interval.

    Pass offer_id to read the bundled synthetic exposures, or pass control
    and treatment arrays. Bootstrap intervals report progress and stop
    between batches if the client cancels. CUPED uses a pre-period
    covariate when one is available. Reported arm means stay on the
    original scale; the lift uses adjusted outcomes when CUPED is on.
check_srmA

Chi-square test of observed arm counts against intended weights.

A mismatch means the counts are unlikely given the weights. It does not identify the cause.

list_experimentsC

List experiment records stored in this server's SQLite file.

create_experimentC

Store a running experiment. This does not assign traffic by itself.

close_experimentA

Mark an experiment closed after the user confirms.

The confirmation form is elicited when the client supports it. If the client does not, the experiment stays open until this tool is called again with confirm=true. Declining or cancelling the form does not close it.

draft_readoutA

Ask the client model to summarize an experiment, with a template fallback.

Sampling is deprecated in the 2026-07-28 spec and many clients do not offer it. Those clients get a deterministic draft built from the same facts.

Prompts

Interactive templates invoked by user choice

NameDescription
design_experimentPlan a retail-offer test before any readout is written.
write_readoutOutline a readout that stays inside the computed facts.

Resources

Contextual data attached and managed by the client

NameDescription
dataset_schemaSynthetic dataset column schema.
dataset_summaryRow counts for the bundled synthetic dataset.
experiment_indexAll experiment records in the local store.
Deploy Server
One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

TDQS

A3.6/5.0

Scored across 8 tools

Disambiguation5/5

Each tool targets a distinct phase of the experiment workflow: power planning, assignment, lift analysis, SRM checking, record lifecycle, and readout drafting. The descriptions explicitly clarify boundary cases, such as assign_cells not writing a record and create_experiment not assigning traffic, which further reduces misselection.

Naming Consistency4/5

All tool names use consistent snake_case, and most follow a verb_noun pattern such as assign_cells, analyze_lift, and create_experiment. The main minor deviation is power_calc, which uses a noun abbreviation rather than a clear verb form like calculate_power.

Tool Count5/5

Eight tools are well-scoped for an experiment management and analysis server. Each tool covers a meaningful part of the workflow without obvious redundancy or inflated surface area.

Completeness4/5

The surface covers experimental design, assignment, analysis, SRM checks, record creation/listing/closure, and readout drafting. Minor gaps remain, such as no get_experiment or update_experiment operation, but the core lifecycle is well represented.

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources