reparatio-mcp

Alpha software. Interfaces, tool names, and behaviour may change without notice between versions.

MCP server for Reparatio — inspect, convert, merge, append, and query local data files directly from your AI assistant.

Supports CSV, TSV, Excel (.xlsx / .xls), ODS, JSON, JSONL, GeoJSON, Parquet, Feather, Arrow, ORC, Avro, SQLite, YAML, BSON, SRT, VTT, HTML, Markdown, XML, SQL dumps, PDF (text layer), and files compressed with gzip, bzip2, zstd, or zip.

See also: reparatio-cli (command-line tool) · reparatio-sdk (Python SDK)

How it works

The MCP server runs locally on your machine. When your AI assistant calls a tool, the server reads the file from your disk, sends it to the Reparatio API for processing, and writes the result back to disk — returning the output path to the assistant. Binary data never passes through the MCP protocol layer; only file paths and options travel as JSON.

Related MCP server: KnowledgeMCP

Prerequisites

• Python 3.11 or later — check with python3 --version

• A Reparatio API key — required for convert, merge, append, and query. inspect_file works without a key.

  • Get a key at reparatio.app (Professional plan — $79/mo)

  • The MCP server requires the Professional plan; Standard ($29/mo) does not include API or MCP access

  • Keys are prefixed with rp_

uvx (from uv) is the recommended way to run the server — no install step required. Install uv with:

curl -LsSf https://astral.sh/uv/install.sh | sh

Installation

Option A — run without installing (recommended):

uvx reparatio-mcp

Option B — install permanently:

pip install reparatio-mcp

Client configuration

MCP is a client-side protocol. Your AI assistant application needs to support MCP tool calls. The sections below cover every major client. The LLM backend (Claude, GPT-4o, Grok, DeepSeek, Gemini, a local Ollama model, etc.) is configured separately inside each client — the MCP server works with any of them.

Claude Desktop

Edit claude_desktop_config.json:

{
  "mcpServers": {
    "reparatio": {
      "command": "uvx",
      "args": ["reparatio-mcp"],
      "env": {
        "REPARATIO_API_KEY": "rp_YOUR_KEY"
      }
    }
  }
}

Restart Claude Desktop after saving. You should see "reparatio" listed under MCP servers in the app.

Cursor

Cursor supports MCP servers globally or per-project.

Global (~/.cursor/mcp.json):

{
  "mcpServers": {
    "reparatio": {
      "command": "uvx",
      "args": ["reparatio-mcp"],
      "env": {
        "REPARATIO_API_KEY": "rp_YOUR_KEY"
      }
    }
  }
}

Per-project (.cursor/mcp.json in your project root) — same format.

Reload the MCP server list in Cursor Settings → MCP after saving.

Windsurf

Edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "reparatio": {
      "command": "uvx",
      "args": ["reparatio-mcp"],
      "env": {
        "REPARATIO_API_KEY": "rp_YOUR_KEY"
      }
    }
  }
}

VS Code — GitHub Copilot

VS Code 1.99+ supports MCP natively. Add to your settings.json (Ctrl+Shift+P → "Open User Settings (JSON)"):

{
  "mcp": {
    "servers": {
      "reparatio": {
        "type": "stdio",
        "command": "uvx",
        "args": ["reparatio-mcp"],
        "env": {
          "REPARATIO_API_KEY": "rp_YOUR_KEY"
        }
      }
    }
  }
}

VS Code — Continue

Continue is a VS Code / JetBrains extension that supports many LLM backends (OpenAI, Anthropic, DeepSeek, Grok, Gemini, Ollama, and more) alongside MCP tool servers.

Edit ~/.continue/config.json:

{
  "mcpServers": [
    {
      "name": "reparatio",
      "command": "uvx",
      "args": ["reparatio-mcp"],
      "env": {
        "REPARATIO_API_KEY": "rp_YOUR_KEY"
      }
    }
  ]
}

Continue handles the LLM connection separately. Reparatio tools are available regardless of which model you point Continue at.

Zed

Edit ~/.config/zed/settings.json:

{
  "context_servers": {
    "reparatio": {
      "command": {
        "path": "uvx",
        "args": ["reparatio-mcp"],
        "env": {
          "REPARATIO_API_KEY": "rp_YOUR_KEY"
        }
      }
    }
  }
}

Open WebUI (Ollama and others)

Open WebUI is a self-hosted chat interface that works with Ollama, OpenAI-compatible APIs, and others, and supports MCP tool servers.

1. Go to Settings → Tools → MCP Servers

2. Add a new server:

   • Command: uvx reparatio-mcp

   • Environment: REPARATIO_API_KEY=rp_YOUR_KEY

3. Save and reload.

The tools will be available to whichever model you have selected — Llama 3, Mistral, Qwen, DeepSeek-R1, or any other Ollama model.

Claude Code (CLI)

Add to your project's .mcp.json or run inline:

{
  "mcpServers": {
    "reparatio": {
      "command": "uvx",
      "args": ["reparatio-mcp"],
      "env": {
        "REPARATIO_API_KEY": "rp_YOUR_KEY"
      }
    }
  }
}

Available tools

inspect_file

Detect encoding, count rows, list column types and statistics, and return a data preview. No API key required.

convert_file

Convert a file from any supported input format to any supported output format. Professional plan required ($79/mo).

merge_files

Merge two files using a SQL-style join or append (row-stacking). Professional plan required ($79/mo).

Operations:

append_files

Stack rows from two or more files vertically. Handles column mismatches gracefully. Professional plan required ($79/mo).

query_file

Run a SQL query against a file. The file is loaded as a table named data. Professional plan required ($79/mo).

Supports SELECT, WHERE, GROUP BY, ORDER BY, LIMIT, aggregations, and most scalar functions.

Supported formats

Example prompts

Once the server is connected, you can use natural language:

Inspect ~/data/sales.csv and tell me about the schema.

Convert ~/data/sales.csv to Parquet.

Convert ~/data/report.xlsx, sheet "Q3", to CSV, selecting only the
"date", "region", and "revenue" columns.

Merge orders.csv and customers.csv on customer_id using a left join,
output as Parquet.

Append all the CSV files in ~/data/monthly/ into one file and save it
as ~/data/all_months.parquet.

Query ~/data/events.parquet: give me total revenue by region for 2025,
sorted descending, as a JSON file.

I have ~/data/legacy.csv from an old Windows system — fix the encoding
and convert it to Excel.

Convert ~/data/mainframe.dat (EBCDIC cp037 encoding) to CSV.
Use encoding_override: cp037.

Output paths

If you don't specify output_path, the server picks a sensible default in the same directory as the input:

Pricing

The Professional plan is required for MCP tool access (convert, merge, append, query). inspect_file is free for all users but output is truncated for non-Professional accounts. A Professional plan removes all row limits and enables all tools.

Privacy

Your files are sent to the Reparatio API at reparatio.app for processing. Files are handled in memory and never stored — see the Privacy Policy. If you need to keep data fully local, this tool is not the right fit.

Environment variables

Troubleshooting

"REPARATIO_API_KEY is not set" The key is missing from the env block in your client config. Check spelling — it must be exactly REPARATIO_API_KEY.

"uvx: command not found" Install uv: curl -LsSf https://astral.sh/uv/install.sh | sh, then restart your terminal and your AI client.

"Insufficient plan" The tool you called requires a Professional plan key ($79/mo). inspect_file is always free.

"File not found" Use an absolute path (starting with / or ~). Relative paths are resolved from the server process's working directory, which may not be what you expect.

"Parse failure" The file could not be read in the detected format. Try setting fix_encoding: true, specifying a delimiter, or confirming the file extension matches the actual format.

Tools not appearing in the client Restart the client after editing the config. Check that uvx is on the PATH visible to the client process — on macOS/Linux, confirm with which uvx.