# Autocarver

> Autocarver is an open-source MCP server for analytics and data. Local MCP server that qualifies dataset columns and carves them into bins for risk models. It runs locally on your machine and is started by your AI client.

- **Kind:** MCP server
- **Pricing:** Open source
- **Platforms:** CLI, Linux, macOS, Windows
- **Website:** https://autocarver.readthedocs.io/en/stable/mcp.html
- **Repository:** https://github.com/mdefrance/autocarver
- **MCP Registry name:** io.github.mdefrance/autocarver
- **Categories:** analytics and data, developer tools
- **Listing updated:** 2026-11-14

## About

AutoCarver is a Python library that turns numeric, categorical and ordinal columns into supervised, human-readable bins, aimed at credit scoring, fraud detection and risk modeling. Its MCP server lets an assistant drive the first judgment-heavy steps from a file path: decide which columns are numerical, categorical, ordinal, datetime or hierarchies, then carve them against a target.

A typical session chains load_dataset, inspection helpers such as list_columns and feature_distribution, drafting with suggest_features and set_feature, then run_carver and save_carver. Inspection tools return summaries rather than raw columns. The scope stops at the carved-feature artifact, a saved JSON; it does not train a downstream model.

The docs say the server runs entirely on your machine and that only what your own LLM client shares leaves it. They also warn that qualification is only as good as the model, so a human should confirm feature definitions before production use. Install with pip install "autocarver[mcp]".

## Features

- **Column qualification:** Suggests whether each column is numerical, categorical, ordinal, datetime or nested.
- **Summary-only inspection:** Read-only tools return profiles, not raw data, to fit a model's context.
- **Carving against a target:** run_carver bins features for binary, multiclass, ordinal or continuous targets.
- **Saved carver:** save_carver writes a JSON that restores both carver and features.
- **Runs locally:** Reads .csv or .parquet files on your machine over stdio.
- **Notebook path:** The same session logic can be used with your own LLM client in a notebook.

## Good for

- Qualify the columns of a loan dataset before scorecard work
- Bin features against a default flag
- Check hierarchy nesting between categorical columns
- Save a carver for reuse in a modelling pipeline
- Get a first draft of feature types for human review

## Install

### Claude Code (Terminal)

```sh
claude mcp add autocarver -- uvx autocarver
```

### Claude Desktop (claude_desktop_config.json)

```json
{
  "mcpServers": {
    "autocarver": {
      "command": "uvx",
      "args": [
        "autocarver"
      ]
    }
  }
}
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "autocarver": {
      "command": "uvx",
      "args": [
        "autocarver"
      ]
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "autocarver": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "autocarver"
      ]
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "autocarver": {
      "command": "uvx",
      "args": [
        "autocarver"
      ]
    }
  }
}
```

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "autocarver": {
      "command": "uvx",
      "args": [
        "autocarver"
      ]
    }
  }
}
```

## Tools

- `load_dataset`: Load a CSV or Parquet file as the working dataset.
- `list_columns`: Show dtype, cardinality, missingness and suggested kind per column.
- `profile_column`: Profile one column.
- `feature_distribution`: Show a feature's distribution.
- `validate_nesting`: Check nesting between columns.
- `suggest_features`: Draft feature definitions.
- `set_feature`: Set a feature's type.
- `run_carver`: Carve features against the target.
- `save_carver`: Save the carver to JSON.

## FAQ

### Does my dataset leave my machine?

The docs say the server runs fully locally; only what your own LLM client shares with its provider leaves it.

### Which files can it load?

CSV and Parquet files.

### Does it train a model?

No. The workflow ends with the carved-feature artifact.

### Which clients are mentioned?

VS Code Copilot, Claude Desktop and Cursor.

### Is Autocarver free?

Yes. Autocarver is open source, so you can use and self-host it for free.

### How do I add Autocarver to Claude Code?

Run this command in your terminal: claude mcp add autocarver -- uvx autocarver

### What are the best alternatives to Autocarver?

The closest alternatives to Autocarver are REA, Codebase Memory, GitHub, Serena MCP: the IDE for your agent and XERJ.

## Alternatives

- [REA](https://ai-powered.com/mcp/rea): REA is an open-source MCP server for developer tools.
- [Codebase Memory](https://ai-powered.com/mcp/codebase-memory): Codebase Memory is an open-source MCP server for knowledge and memory.
- [GitHub](https://ai-powered.com/mcp/github-2): GitHub is an open-source MCP server for developer tools.
- [Serena MCP: the IDE for your agent](https://ai-powered.com/mcp/serena-mcp-the-ide-for-your-agent): Serena MCP: the IDE for your agent is an open-source MCP server for developer tools.
- [XERJ](https://ai-powered.com/mcp/xerj): XERJ is an open-source MCP server for knowledge and memory.
- [Agent-Native Analytics](https://ai-powered.com/mcp/agent-native-analytics): Agent-Native Analytics is an open-source MCP server for analytics and data.
- [Ouroboros](https://ai-powered.com/mcp/ouroboros): Ouroboros is an open-source MCP server for developer tools.
- [Claude Code Ultimate Guide](https://ai-powered.com/mcp/claude-code-ultimate-guide): Claude Code Ultimate Guide is an open-source MCP server for ai models and llm tooling.

---
Source: https://ai-powered.com/mcp/autocarver
