# leap-laboratories/discovery-engine [Health: Active]

**Category:** 🧮 Data Science Tools  
**Repository:** https://github.com/leap-laboratories/discovery-engine  
**GitHub Stars:** 7  
**Views:** 4  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/leap-laboratories-discovery-engine

## Description
Superhuman exploratory data analysis that finds the feature interactions and subgroup effects that LLMs and manual exploration miss — with p-values, effect sizes, and literature citations. Data goes in, validated insights come out. Free for public data.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "discovery-engine": {
    "command": "uvx",
    "args": ["discovery-engine-api"],
    "env": {
      "DISCOVERY_API_KEY": ""
    }
  }
}
```

**Requires environment variables:** `DISCOVERY_API_KEY` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What leap-laboratories/discovery-engine MCP server does

The leap-laboratories/discovery-engine MCP server exposes Disco's exploratory data analysis workflow through MCP. It accepts a tabular dataset and a target column, then searches for combinations of feature conditions associated with meaningful changes in the target. This can include interactions between continuous variables, subgroup effects, and conditional relationships that are not obvious from examining individual columns.

Each returned pattern contains structured evidence. Depending on the result, that evidence can include the matching conditions, target-change direction, absolute target change, support count, support percentage, p-value, novelty classification, explanation of the novelty assessment, and literature citations. The service also provides a generated summary and a URL for an interactive report.

The tool is intended for discovery-oriented analysis. It is not presented as a replacement for pandas-based summary statistics, visualization, filtering, or SQL querying.

## How it works

Disco first trains machine-learning models on part of the supplied data and uses interpretability methods to extract candidate patterns. It then evaluates those patterns against held-out data and applies Benjamini-Hochberg false-discovery-rate correction. Patterns that survive validation are checked against academic literature through semantic search, allowing results to be classified as novel or confirmatory.

Analyses can take several minutes. The hosted service reports progress such as queue position, estimated wait, current pipeline step, and estimated completion time. Agent workflows can submit work and check its status separately rather than waiting for the analysis request to finish.

Data preparation affects the usefulness of the output. Identifiers, target leakage, and columns that are alternative encodings or components of the target should be excluded, because they can produce statistically real but tautological findings.

## Setup and configuration

The leap-laboratories/discovery-engine MCP server is hosted at `https://disco.leap-labs.com/mcp`, so no local MCP package installation is required. Configure an MCP client with that URL and provide `DISCOVERY_API_KEY` in the server environment.

An API key can be obtained through the service's signup and email-verification endpoints or through its developer page. The MCP tool list also includes signup and login verification operations. Public analyses are free but publish the data and report; private analyses use credits. The analysis request can specify visibility, analysis depth, column descriptions, excluded columns, a title, a description, and whether to use LLM-assisted processing.

## Tools and capabilities

The available MCP tools are:

- `discovery_list_plans` and `discovery_estimate` for plan and cost-related information.
- `discovery_upload` for sending data to the service.
- `discovery_analyze` for starting an analysis.
- `discovery_status` for checking a submitted run.
- `discovery_get_results` for retrieving completed results.
- `discovery_account` for account information.
- `discovery_signup`, `discovery_signup_verify`, `discovery_login`, and `discovery_login_verify` for account access flows.
- `discovery_add_payment_method`, `discovery_subscribe`, and `discovery_purchase_credits` for paid usage management.

Use the server to investigate which variables and combinations relate to a target, identify candidate subgroup effects, compare findings with published research, or produce structured evidence for a downstream agent workflow.

## Limitations and notes

The service does not guarantee that a statistically validated pattern is causal. Input columns that encode the target or contain leakage can lead to misleading results, so data preparation is required. Public runs make the data and report public, while private runs consume credits. Analyses may take a few minutes, and LLM-assisted processing is slower and more expensive than the default mode.

_Full upstream README: https://allmcps.com/mcp/leap-laboratories-discovery-engine/readme_