llm-tools/mcps/dicom_mcp/INSTALL.md

# Installation Guide

This guide covers installing the DICOM MCP server for use with Claude Desktop, Claude Code, or any other MCP-compatible client.

## Table of Contents

- [Supported Clients](#supported-clients)
- [1. Installation](#1-installation)
  - [1a. Standalone Package (Non-Developers)](#1a-standalone-package-non-developers)
  - [1b. Developer Installation (Poetry)](#1b-developer-installation-poetry)
- [2. Client Configuration](#2-client-configuration)
  - [2a. Claude Desktop](#2a-claude-desktop)
  - [2b. Claude Code](#2b-claude-code)
- [3. PII Filtering](#3-pii-filtering)
- [4. Environment Variables](#4-environment-variables)
- [5. Verifying the Installation](#5-verifying-the-installation)
- [6. Troubleshooting](#6-troubleshooting)

---

## Supported Clients

This MCP server communicates via **stdio** — it runs as a local subprocess on your machine. This means it only works with Claude clients that can spawn local processes.

| Client | Supported? | Config file | Notes |
|--------|-----------|-------------|-------|
| **Claude Desktop** (native app) | Yes | `claude_desktop_config.json` | GUI-based, single config file per platform |
| **Claude Code** (terminal) | Yes | `.mcp.json` or `~/.claude.json` | CLI-based, multiple scope levels |
| **Claude Code** (VS Code / IDE) | Yes | Same as above | Same CLI, same config files — the IDE terminal makes no difference |
| **claude.ai** (web browser) | No | N/A | Cannot spawn local processes; would require an HTTP/SSE remote server wrapper |

This server implements the open [Model Context Protocol](https://modelcontextprotocol.io/) standard, so it also works with other MCP-compatible clients such as **ChatGPT**, Cursor, and Windsurf. Configuration will vary by client — refer to your client's MCP documentation. The instructions below cover Claude Desktop and Claude Code specifically.

If you're unsure which client you have: **Claude Desktop** is the native app with a chat window. **Claude Code** is the `claude` command you run in a terminal (standalone, VS Code, or any other IDE). **claude.ai** is the website at claude.ai.

---

## 1. Installation

Choose the installation method that matches your environment.

### 1a. Standalone Package (Non-Developers)

**No Python installation required.** Ideal for radiographers, QA analysts, and other team members who don't have a development environment. Everything is self-contained — Python, pydicom, numpy, and all dependencies are bundled.

#### macOS

1. Download the standalone package for your Mac:
   - Apple Silicon (M1/M2/M3/M4): `dicom_mcp_standalone_arm64.tar.gz`
   - Intel Mac: `dicom_mcp_standalone_x86_64.tar.gz`

2. Extract and install:
```bash
tar -xzf dicom_mcp_standalone_arm64.tar.gz
cd dicom_mcp_standalone_arm64
./install_to_claude.sh
```

3. Restart Claude Desktop.

The installer configures Claude Desktop automatically.

#### Linux

1. Download `dicom_mcp_standalone_linux_x86_64.tar.gz` (or `aarch64` for ARM).

2. Extract and install:
```bash
tar -xzf dicom_mcp_standalone_linux_x86_64.tar.gz
cd dicom_mcp_standalone_linux_x86_64
./install_to_claude.sh
```

3. Restart your MCP client.

### 1b. Developer Installation (Poetry)

For developers who want to modify the code, run tests, or extend the server.

#### Prerequisites

- Python 3.12 or higher
- [Poetry](https://python-poetry.org/docs/#installation) (Python package manager)

#### Quick Setup

```bash
# Clone the repository
git clone <repository-url>
cd dicom_mcp

# Run the automated installer (macOS, configures Claude Desktop)
./install.sh
```

#### Manual Setup

1. Install dependencies:
```bash
poetry install --with dev
```

2. Verify the installation:
```bash
poetry run pytest -v --tb=short
```
You should see all 138 tests passing.

3. Start the server (to test it manually):
```bash
poetry run python -m dicom_mcp
```

After installation, proceed to [Client Configuration](#2-client-configuration) below.

---

## 2. Client Configuration

Configuration is different for Claude Desktop and Claude Code. If you used the standalone installer (`install_to_claude.sh`), Claude Desktop is already configured — skip to [Verifying the Installation](#5-verifying-the-installation).

### 2a. Claude Desktop

Claude Desktop reads its MCP server configuration from a single JSON file. The location depends on your OS:

| OS | Config file path |
|----|-----------------|
| macOS | `~/Library/Application Support/Claude/claude_desktop_config.json` |
| Linux | `~/.config/Claude/claude_desktop_config.json` |
| Windows | `%APPDATA%\Claude\claude_desktop_config.json` |

#### Using the GUI

1. Open Claude Desktop and go to Settings (gear icon).
2. Navigate to the Developer section.

   <img src="img/settings_developer_option.png" width="150px" alt="Developer settings">

3. Click **Edit Config** or **Add MCP Server**.

   <img src="img/mcp_servers_panel.png" width="450px" alt="MCP servers panel">

4. Add the server configuration using one of the JSON examples below.

#### Poetry (Recommended for Developers)

```json
{
  "mcpServers": {
    "dicom_mcp": {
      "command": "poetry",
      "args": ["run", "python", "-m", "dicom_mcp"],
      "cwd": "/absolute/path/to/dicom_mcp"
    }
  }
}
```

Replace `/absolute/path/to/dicom_mcp` with the actual directory where the project lives.

#### Virtualenv Python Directly

If Claude Desktop has trouble finding Poetry, you can point it at the virtualenv Python directly. First, get the path:

```bash
poetry env info --path
```

Then configure:

```json
{
  "mcpServers": {
    "dicom_mcp": {
      "command": "/path/to/poetry/virtualenv/bin/python",
      "args": ["-m", "dicom_mcp"],
      "cwd": "/absolute/path/to/dicom_mcp"
    }
  }
}
```

After adding the configuration, **restart Claude Desktop** to load the server.

### 2b. Claude Code

Claude Code supports multiple configuration scopes for MCP servers, each stored in a dedicated file (separate from Claude Code's settings files). Choose the scope that matches your use case.

#### Scope Overview

| Scope | Config file | Shared via git? | Use case |
|-------|-------------|-----------------|----------|
| **Project** | `.mcp.json` (project root) | Yes | Team-shared — everyone on the project gets the server |
| **User** | `~/.claude.json` | No | Personal — available in all your projects |
| **Local** | `~/.claude.json` (project-specific section) | No | Personal — available only in the current project |
| **Managed** | System directory (see below) | Enterprise IT | Organisation-wide enforcement |

**Precedence** (highest to lowest): Managed > Local > Project > User. If the same MCP server name appears at multiple scopes, the higher-precedence scope wins.

#### Project Scope (Team-Shared)

Create `.mcp.json` in your project root and commit it to version control. All collaborators will get the server automatically:

```json
{
  "mcpServers": {
    "dicom_mcp": {
      "command": "poetry",
      "args": ["run", "python", "-m", "dicom_mcp"],
      "cwd": "/absolute/path/to/dicom_mcp"
    }
  }
}
```

Replace `/absolute/path/to/dicom_mcp` with the actual directory where the DICOM MCP project lives.

**Note:** Project-scoped servers require user approval on first use. Users can reset approval with `claude mcp reset-project-choices`.

#### User Scope (Personal, All Projects)

Add to `~/.claude.json` to make the server available in every Claude Code session:

```json
{
  "mcpServers": {
    "dicom_mcp": {
      "command": "poetry",
      "args": ["run", "python", "-m", "dicom_mcp"],
      "cwd": "/absolute/path/to/dicom_mcp"
    }
  }
}
```

#### Local Scope (Personal, One Project)

The local scope is stored in `~/.claude.json` but scoped to a specific project directory. The easiest way to add a local-scoped server is via the CLI:

```bash
claude mcp add dicom_mcp \
  --scope local \
  -- poetry run python -m dicom_mcp
```

#### Managed Scope (Enterprise)

For organisation-wide deployment, IT administrators can place a managed configuration in:

| OS | Path |
|----|------|
| macOS | `/Library/Application Support/ClaudeCode/managed-mcp.json` |
| Linux/WSL | `/etc/claude-code/managed-mcp.json` |
| Windows | `C:\Program Files\ClaudeCode\managed-mcp.json` |

Managed servers cannot be overridden by users and take the highest precedence.

#### Using the CLI

Claude Code provides a CLI for managing MCP servers without editing JSON files directly:

```bash
# Add a server (default: local scope)
claude mcp add dicom_mcp -- poetry run python -m dicom_mcp

# Add at a specific scope
claude mcp add dicom_mcp --scope user -- poetry run python -m dicom_mcp
claude mcp add dicom_mcp --scope project -- poetry run python -m dicom_mcp

# List all configured servers
claude mcp list

# Get details for a specific server
claude mcp get dicom_mcp

# Remove a server
claude mcp remove dicom_mcp
```

#### Virtualenv Python Fallback

If Claude Code has trouble finding Poetry, point directly at the virtualenv Python. First, get the path:

```bash
poetry env info --path
```

Then use the full Python path in your configuration (at any scope):

```json
{
  "mcpServers": {
    "dicom_mcp": {
      "command": "/path/to/poetry/virtualenv/bin/python",
      "args": ["-m", "dicom_mcp"],
      "cwd": "/absolute/path/to/dicom_mcp"
    }
  }
}
```

After saving any configuration, restart Claude Code (or start a new session) for the changes to take effect. You can verify the server is loaded by running `/mcp` in the Claude Code interface.

---

## 3. PII Filtering

Patient-identifying tags can be redacted from all tool output by setting an environment variable. This works with any client and any configuration method.

Add an `env` block to your MCP server configuration:

```json
{
  "mcpServers": {
    "dicom_mcp": {
      "command": "...",
      "args": ["..."],
      "env": {
        "DICOM_MCP_PII_FILTER": "true"
      }
    }
  }
}
```

When enabled, the following tags are replaced with `[REDACTED]` in all tool output:
- PatientName
- PatientID
- PatientBirthDate
- PatientSex

This affects `dicom_get_metadata`, `dicom_compare_headers`, `dicom_summarize_directory`, and `dicom_query`. All other tools do not expose patient data and are unaffected.

**Standalone users:** Edit the Claude Desktop config file that `install_to_claude.sh` created (see [Claude Desktop config locations](#2a-claude-desktop)) and add the `env` block to the existing `dicom_mcp` entry. Keep the `command` value the installer wrote — you're only adding the `env` section.

To disable, unset the variable or set it to any value other than `true`, `1`, or `yes`.

---

## 4. Environment Variables

| Variable | Default | Description |
|----------|---------|-------------|
| `DICOM_MCP_PII_FILTER` | `false` | Set to `true`, `1`, or `yes` to redact patient tags in tool output |
| `DICOM_MCP_MAX_FILES` | `1000` | Maximum number of DICOM files to scan per directory operation |

---

## 5. Verifying the Installation

Once configured, restart your client and verify the server is working:

1. **Claude Desktop**: Ask Claude "What MCP tools do you have available?" — you should see 17 DICOM tools listed.

2. **Claude Code**: The tools will appear in the tool list. You can also test directly:
   ```
   List your available DICOM MCP tools
   ```

3. **MCP Inspector** (for debugging):
   ```bash
   npx @modelcontextprotocol/inspector poetry run python -m dicom_mcp
   ```

---

## 6. Troubleshooting

### Server not appearing after configuration

- Double-check that the `cwd` path is correct and contains the `dicom_mcp/` package directory.
- Ensure Poetry is installed and on the PATH, or use the virtualenv Python directly.
- Restart the client after any configuration change.

### ModuleNotFoundError

The configured Python environment is missing dependencies. Either:
- Run `poetry install` in the project directory, or
- Switch to the virtualenv Python configuration (see above).

### Permission denied on standalone launcher

```bash
chmod +x run_dicom_mcp.sh install_to_claude.sh
```

### Poetry not found by Claude Desktop

Claude Desktop may not inherit your shell's PATH. Use the virtualenv Python directly instead of the Poetry command (see configuration examples above).