zsv-mcp

A Go-based Model Context Protocol (MCP) server that exposes tools for the zsv CLI (see: https://github.com/liquidaty/zsv).

This project uses the official MCP Go SDK: github.com/modelcontextprotocol/go-sdk.

Features

• zsv_run: run zsv with explicit argument arrays and return captured stdout.
• zsv_help: run zsv --help and return captured stdout.
• Configurable runtime values through environment variables.
• Unit tests for core tool behavior.

Repository Layout

.
|- .cursor/
|  |- mcp.json             # Cursor MCP config
|- .github/
|  |- copilot-instructions.md
|- .mcp.json               # Claude Code MCP config
|- .vscode/
|  |- mcp.json             # VS Code / GitHub Copilot MCP config
|  |- settings.json        # VS Code MCP config (legacy format)
|- CLAUDE.md               # Cursor/Claude AI instructions for using zsv MCP tools
|- config/
|  |- config.go            # Environment-based configuration
|- docs/
|  |- TOOL_REFERENCE.md    # detailed tool contracts and examples
|- main.go                 # MCP server entry point
|- main_test.go            # Core handler tests
|- prompts/
|  |- tool_usage.go        # tool usage guidance prompt
|- resources/
|  |- tool_usage.go        # tool usage guidance resource
|- sample_data/
|  |- data.csv             # sample CSV for example prompts
|- SETUP.md                # VS Code local setup notes
|- tools/
|  |- help.go              # zsv_help tool
|  |- run.go               # zsv_run tool
|  |- run_help_test.go     # zsv_run/zsv_help tool tests

Requirements

• Go 1.23+
• zsv CLI available on your PATH for zsv_run and zsv_help tools, or ZSV_PATH set to the binary location - see Installation details.

Quick Start

1. Clone and enter the project directory.

2. Install dependencies:

go mod download

3. Build:

# Linux/macOS
go build -o zsv-mcp

# Windows
go build -o zsv-mcp.exe

4. Start the server:

# Linux/macOS
./zsv-mcp

# Windows
zsv-mcp.exe

The server communicates over stdio. See MCP Client Configuration Files to connect it to your editor.

5. Run tests:

go test ./...

Configuration

Environment variables are read in config/config.go.

• SERVER_NAME (default: zsv-mcp)
• VERSION (default: v1.0.0)
• LOG_LEVEL (default: info)
• ZSV_PATH (default: zsv) used by zsv_run and zsv_help

Logging notes:

• The server writes logs to stderr only to avoid interfering with MCP stdio protocol traffic on stdout.
• Supported LOG_LEVEL values: debug, info, warn (or warning), error.
• Unknown LOG_LEVEL values fall back to info and emit a warning at startup.

PowerShell example:

$env:SERVER_NAME = "zsv-mcp"
$env:VERSION = "v1.0.0"
$env:LOG_LEVEL = "debug"
$env:ZSV_PATH = "zsv"
zsv-mcp.exe

MCP Client Configuration Files

The repository ships with pre-configured files so MCP-compatible clients can discover and launch the server without manual setup. All configs expect the binary to be named zsv-mcp (no .exe extension) — see the cross-platform note in Quick Start above.

.mcp.json

Project-level MCP discovery file read by Claude Code. It registers zsv-mcp as a stdio server with no extra arguments or environment overrides. The command path is relative to the project root, so the client must be launched from that directory (or the path adjusted accordingly).

{
  "mcpServers": {
    "zsv-mcp": {
      "type": "stdio",
      "command": "./zsv-mcp",
      "args": [],
      "env": {
        "ZSV_PATH": "zsv",
        "LOG_LEVEL": "debug"
      }
    }
  }
}

.vscode/mcp.json

VS Code-specific MCP server configuration introduced in newer VS Code releases. VS Code reads this file to populate the Chat MCP tool list. It uses ${workspaceFolder} so the path resolves correctly regardless of where VS Code is opened. GitHub Copilot in VS Code also uses this file.

{
  "servers": {
    "zsv-mcp": {
      "command": "${workspaceFolder}/zsv-mcp",
      "type": "stdio",
      "args": [],
      "env": {
        "ZSV_PATH": "zsv",
        "LOG_LEVEL": "debug"
    }
  }
}

.vscode/settings.json

Workspace settings file that registers the server under the mcpServers key. Earlier versions of VS Code's MCP integration use this file rather than .vscode/mcp.json. The disabled field controls whether the server is active, and alwaysAllow lists any tool calls that are approved automatically without a user prompt.

{
  "mcpServers": {
    "zsv-mcp": {
      "command": "${workspaceFolder}/zsv-mcp",
      "args": [],
      "disabled": false,
      "alwaysAllow": [],
      "env": {
        "ZSV_PATH": "zsv",
        "LOG_LEVEL": "debug"
      }
    }
  }
}

Both .vscode/mcp.json and .vscode/settings.json point to the same binary; which file VS Code consults depends on its version, so both are included for compatibility.

.cursor/mcp.json

Cursor MCP configuration. Cursor's support for ${workspaceFolder} is inconsistent across versions, so this config references the binary by name only. Ensure zsv-mcp is on your PATH (for example, via go install) or replace the command value with the absolute path to the binary.

{
  "mcpServers": {
    "zsv-mcp": {
      "command": "zsv-mcp",
      "args": [],
      "env": {
        "ZSV_PATH": "zsv",
        "LOG_LEVEL": "debug"
      }
    }
  }
}

Claude Desktop

Claude Desktop uses a user-level config file, not a repo-level one. Add the following to your claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "zsv-mcp": {
      "command": "/absolute/path/to/zsv-mcp",
      "args": [],
      "env": {
        "ZSV_PATH": "zsv",
        "LOG_LEVEL": "debug"
      }
    }
  }
}

Tool Summary

See TOOL_REFERENCE.md for complete request and response details.

• zsv_run(cmd[])
• zsv_help()

Example Prompts

The following prompts can be used with any MCP-compatible AI client (Claude, Cursor, VS Code Copilot, etc.) to exercise the zsv MCP tools. Most prompts explicitly name the MCP tool to use so that the AI does not scan project source files for context. The last one does not name the tool(s) explicitly and instead relies on the AI client deducing them from the MCP server and the guidance provided for it.

For verification purposes, fresh prompts should be used for each one to ensure that the AI client is not influenced by prior prompt message results. During processing, you may see the AI client self-correct as it learns how to issue well-formed requests to the MCP server.

Verify installation

Using the zsv_help MCP tool, show me the zsv help text.

The AI calls zsv_help() and returns the zsv CLI's top-level help output, listing available subcommands. If this fails, either the zsv MCP tooling is not registered, zsv is not on PATH or ZSV_PATH is not set correctly.

Select columns from a file

Using the zsv_run MCP tool, show me only the Name and Country columns from sample_data/data.csv.

The AI calls zsv_run with the select subcommand to extract two columns and returns a filtered CSV.

Calculate the average of a column

Using the zsv_run MCP tool, calculate the average age of the people in sample_data/data.csv.

The AI calls zsv_run using the sql subcommand with a query such as select avg(cast([Age] as real)) from data, then reports the result.

Multi-stage operation

Using the zsv_run MCP tool, select only the Name and Age columns from sample_data/data.csv, then sort the result by Age descending.

The AI will decide how it handles this. It may decide to issue two zsv_run calls (select to narrow columns, then sql to sort) or to combine both operations into a single sql query. It will then report the final result.

Multi-stage operation (no specific MCP tool named)

Using zsv, select only the Name and Country columns from sample_data/data.csv, then sort the result by Name in ascending order.

The AI will determine how to handle this, likely consulting the zsv MCP usage guide and ultimately executing a sql call via zsv_run to generate and then display the final result.

Development

• Local editor setup: SETUP.md
• Contribution workflow: CONTRIBUTING.md