rawtherapee-mcp-server
Enables AI-assisted RAW photo development via RawTherapee CLI, with a visual feedback loop that allows the LLM to see and iteratively edit images.
README
RawTherapee MCP Server
Cross-platform Model Context Protocol (MCP) server for AI-assisted RAW photo development via RawTherapee CLI. Provides 49 tools for profile generation, image processing, visual previews, batch operations, device presets, luminance-based local adjustments, lens correction, film simulation LUTs, profile inheritance, and metadata privacy.
What makes it unique: The LLM can see the photos it's editing. Preview tools return inline Base64 images via MCP's ImageContent protocol, creating a visual feedback loop where the AI analyzes the image, adjusts settings, previews the result, and iterates — just like a human editor.
MCP Client Compatibility
Not all MCP clients handle inline images the same way. The visual feedback loop requires a client that renders ImageContent from tool responses and a backing LLM with vision capabilities.
| Client | MCP Support | Image Display | Visual Workflow | Status |
|---|---|---|---|---|
| Claude Desktop | Full | Yes | Full | Tested |
| Claude Code | Full | No (terminal) | Partial | Tested — images not rendered in terminal, but processing and text analysis work fully |
| Cursor | Full | Should work | Should work | Untested — MCP docs indicate ImageContent support |
| Windsurf | Full | Should work | Should work | Untested |
| Cline | Partial | Unknown | Unknown | Untested — community reports suggest ImageContent may not render (#1865) |
| Zed | Full | Unknown | Unknown | Untested |
Minimum requirements for the full visual workflow:
- MCP client renders
ImageContent(type: "image", data: base64, mimeType: "image/jpeg") from tool responses - Backing LLM supports vision/image analysis (e.g. Claude with vision)
- Tool response size accommodates ~150KB previews (most clients: 1MB limit)
Text-only clients: All 49 tools work without inline images. Preview tools return file paths instead. The LLM can still read EXIF metadata, histogram statistics, generate profiles, batch process, and use luminance presets — the visual feedback loop is the only feature that requires image support.
Prerequisites
- Python 3.11+
- RawTherapee 5.9+ with CLI component installed
- An MCP-compatible client (see table above)
Verify the RawTherapee CLI is reachable before proceeding:
Windows (PowerShell):
& "C:\Program Files\RawTherapee\5.12\rawtherapee-cli.exe" --version
macOS:
/Applications/RawTherapee.app/Contents/MacOS/rawtherapee-cli --version
Linux:
rawtherapee-cli --version
Installation
Two supported methods - pick the one that matches your priorities:
| Method | Choose when you want |
|---|---|
| uvx | Minimal setup, no persistent install, standard MCP community pattern |
| pip + venv | Explicit version control, predictable updates, easier debugging, air-gapped environments |
Option A - uvx
No persistent install. uvx downloads, caches, and runs the server on demand.
Important: always pin a specific version. The unpinned form ("rawtherapee-mcp-server") locks to the first resolved version and will not auto-upgrade on its own. See Updating for the full refresh procedure.
<details> <summary><strong>Windows</strong> (<code>%APPDATA%\Claude\claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server@1.0.4"],
"env": {
"RT_CLI_PATH": "C:\\Program Files\\RawTherapee\\5.12\\rawtherapee-cli.exe",
"RT_OUTPUT_DIR": "C:\\Users\\YourName\\Pictures\\rawtherapee-output"
}
}
}
}
</details>
<details> <summary><strong>macOS</strong> (<code>~/Library/Application Support/Claude/claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server@1.0.4"],
"env": {
"RT_OUTPUT_DIR": "/Users/yourname/Pictures/rawtherapee-output"
}
}
}
}
RT CLI is auto-detected at /Applications/RawTherapee.app/Contents/MacOS/rawtherapee-cli. Set RT_CLI_PATH if your installation differs.
</details>
<details> <summary><strong>Linux</strong> (<code>~/.config/Claude/claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server@1.0.4"],
"env": {
"RT_OUTPUT_DIR": "/home/yourname/Pictures/rawtherapee-output"
}
}
}
}
RT CLI is auto-detected at /usr/bin/rawtherapee-cli. Set RT_CLI_PATH if needed.
</details>
Option B - pip + venv
Explicit install with a dedicated virtual environment. You control exactly which version is running and updates are transparent.
1. Create the venv and install
Windows (PowerShell):
python -m venv "$env:USERPROFILE\.rawtherapee-mcp-env"
& "$env:USERPROFILE\.rawtherapee-mcp-env\Scripts\pip.exe" install rawtherapee-mcp-server
macOS / Linux:
python3 -m venv ~/.rawtherapee-mcp-env
~/.rawtherapee-mcp-env/bin/pip install rawtherapee-mcp-server
2. Configure Claude Desktop
Point Claude Desktop at the venv's entry-point script directly. No activation step is needed - the path is absolute.
<details> <summary><strong>Windows</strong> (<code>%APPDATA%\Claude\claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "C:\\Users\\YourName\\.rawtherapee-mcp-env\\Scripts\\rawtherapee-mcp-server.exe",
"args": [],
"env": {
"RT_CLI_PATH": "C:\\Program Files\\RawTherapee\\5.12\\rawtherapee-cli.exe",
"RT_OUTPUT_DIR": "C:\\Users\\YourName\\Pictures\\rawtherapee-output"
}
}
}
}
Replace YourName with your actual Windows username.
</details>
<details> <summary><strong>macOS</strong> (<code>~/Library/Application Support/Claude/claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "/Users/yourname/.rawtherapee-mcp-env/bin/rawtherapee-mcp-server",
"args": [],
"env": {
"RT_OUTPUT_DIR": "/Users/yourname/Pictures/rawtherapee-output"
}
}
}
}
</details>
<details> <summary><strong>Linux</strong> (<code>~/.config/Claude/claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "/home/yourname/.rawtherapee-mcp-env/bin/rawtherapee-mcp-server",
"args": [],
"env": {
"RT_OUTPUT_DIR": "/home/yourname/Pictures/rawtherapee-output"
}
}
}
}
</details>
3. Verify the install
Windows:
& "$env:USERPROFILE\.rawtherapee-mcp-env\Scripts\python.exe" -c "import rawtherapee_mcp; print(rawtherapee_mcp.__version__)"
macOS / Linux:
~/.rawtherapee-mcp-env/bin/python -c "import rawtherapee_mcp; print(rawtherapee_mcp.__version__)"
Expected output: the installed version number (e.g. 1.0.4).
After editing claude_desktop_config.json, fully quit and restart Claude Desktop - closing the window alone leaves a background process running on both macOS and Windows.
Client Configuration
Claude Desktop
The RawTherapee CLI path is auto-detected on most systems. Set RT_CLI_PATH if auto-detection fails or RawTherapee is in a non-standard location.
<details> <summary><strong>Windows</strong> (<code>%APPDATA%\Claude\claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server"],
"env": {
"RT_CLI_PATH": "C:\\Program Files\\RawTherapee\\5.12\\rawtherapee-cli.exe",
"RT_OUTPUT_DIR": "D:\\Photos\\Processed"
}
}
}
}
</details>
<details> <summary><strong>macOS</strong> (<code>~/Library/Application Support/Claude/claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server"],
"env": {
"RT_OUTPUT_DIR": "/Users/you/Pictures/Processed"
}
}
}
}
RT CLI is auto-detected at /Applications/RawTherapee.app/Contents/MacOS/rawtherapee-cli.
</details>
<details> <summary><strong>Linux</strong> (<code>~/.config/Claude/claude_desktop_config.json</code>)</summary>
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server"],
"env": {
"RT_OUTPUT_DIR": "/home/you/Pictures/Processed"
}
}
}
}
RT CLI is auto-detected at /usr/bin/rawtherapee-cli, /usr/local/bin/rawtherapee-cli, or /snap/bin/rawtherapee-cli.
</details>
<details> <summary><strong>Development (from source)</strong></summary>
Use uv to run from the cloned repository:
{
"mcpServers": {
"rawtherapee": {
"command": "uv",
"args": ["--directory", "/path/to/rawtherapee-mcp-server", "run", "rawtherapee-mcp-server"],
"env": {
"RT_OUTPUT_DIR": "/home/you/Pictures/Processed"
}
}
}
}
</details>
Claude Code
# Published package
claude mcp add rawtherapee -- uvx rawtherapee-mcp-server
# Development (from source)
claude mcp add rawtherapee -- uv --directory /path/to/rawtherapee-mcp-server run rawtherapee-mcp
Cursor
Add to .cursor/mcp.json in your project:
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server"],
"env": {
"RT_CLI_PATH": "/usr/bin/rawtherapee-cli",
"RT_OUTPUT_DIR": "/home/you/Pictures/Processed"
}
}
}
}
Windsurf
Add to your Windsurf MCP configuration:
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server"],
"env": {
"RT_CLI_PATH": "/usr/bin/rawtherapee-cli",
"RT_OUTPUT_DIR": "/home/you/Pictures/Processed"
}
}
}
}
Cline (VS Code)
Add to your Cline MCP settings:
{
"mcpServers": {
"rawtherapee": {
"command": "uvx",
"args": ["rawtherapee-mcp-server"],
"env": {
"RT_CLI_PATH": "/usr/bin/rawtherapee-cli",
"RT_OUTPUT_DIR": "/home/you/Pictures/Processed"
}
}
}
}
Note: Cline may not render inline images from tool responses. All text-based tools work normally.
For pip + venv config examples for all clients, see docs/CLIENT_COMPATIBILITY.md.
Quick Start
After installation and client configuration, try this workflow:
- "Analyze this photo" —
analyze_imagereads EXIF, computes histogram, generates a thumbnail, and returns processing recommendations - "Make it warmer with more contrast" —
generate_pp3_profilecreates a profile with warm white balance and contrast boost,preview_before_aftershows the difference - "The shadows are too dark" —
add_luminance_adjustmentadds a shadow recovery spot,preview_with_adjustmentsshows the result - "Export for my phone" —
process_rawwithdevice_presetcrops and processes at the right aspect ratio
Updating
Both installation methods require a two-step procedure. Upgrading the package alone is not enough - Claude Desktop caches the MCP server's tool list. Always follow the full sequence: update package → fully quit Claude Desktop → relaunch.
With uvx
uvx creates a per-package receipt that locks the resolved version. --refresh re-downloads the same locked version; it does not upgrade. To get a new release, clear the receipt and update the version pin:
Windows (PowerShell):
uv cache clean rawtherapee-mcp-server
# Update the version pin in %APPDATA%\Claude\claude_desktop_config.json
# e.g. change @1.0.4 to @1.0.5
Get-Process *claude* -ErrorAction SilentlyContinue | Stop-Process -Force
Start-Process "$env:LOCALAPPDATA\AnthropicClaude\Claude.exe"
macOS:
uv cache clean rawtherapee-mcp-server
# Update the version pin in ~/Library/Application Support/Claude/claude_desktop_config.json
osascript -e 'quit app "Claude"' && sleep 2 && open -a Claude
Linux:
uv cache clean rawtherapee-mcp-server
# Update the version pin in ~/.config/Claude/claude_desktop_config.json
pkill -f claude && sleep 2 && claude &
With pip + venv
Windows:
& "$env:USERPROFILE\.rawtherapee-mcp-env\Scripts\pip.exe" install --upgrade rawtherapee-mcp-server
Get-Process *claude* -ErrorAction SilentlyContinue | Stop-Process -Force
Start-Process "$env:LOCALAPPDATA\AnthropicClaude\Claude.exe"
macOS / Linux:
~/.rawtherapee-mcp-env/bin/pip install --upgrade rawtherapee-mcp-server
# macOS:
osascript -e 'quit app "Claude"' && sleep 2 && open -a Claude
# Linux:
pkill -f claude && sleep 2 && claude &
Verify the update worked
After relaunch, ask Claude:
"What is the current rawtherapee-mcp-server version and how many tools are registered?"
Compare against CHANGELOG.md. If the version string is correct but new tools are missing, see docs/TROUBLESHOOTING.md.
Check GitHub Releases for full changelogs.
Available Tools (49)
Discovery & Configuration (5)
| Tool | Description |
|---|---|
check_rt_status |
Check RawTherapee installation, version, CLI path, and server configuration |
list_templates |
List all available PP3 templates (built-in and custom) |
list_device_presets |
List all device/format crop and resize presets |
list_raw_files |
Scan a directory for supported RAW files |
list_output_files |
List processed output files in the output directory |
Metadata & Analysis (5)
| Tool | Description |
|---|---|
read_exif |
Read EXIF metadata with structured processing recommendations |
analyze_image |
All-in-one analysis: EXIF + histogram + thumbnail + recommendations |
batch_analyze |
Analyze multiple images with EXIF, recommendations, and thumbnails |
get_image_info |
Get dimensions, format, file size with optional inline thumbnail |
get_histogram |
RGB histogram with per-channel statistics, clipping, and SVG visualization |
Profile Management (8)
| Tool | Description |
|---|---|
generate_pp3_profile |
Create a PP3 profile from base template + parameters + device preset |
read_profile |
Display PP3 profile contents in human-readable format |
adjust_profile |
Modify specific parameters in an existing profile |
compare_profiles |
Diff two profiles with optional visual A/B comparison |
save_template |
Save a profile as a reusable custom template |
create_template_from_description |
Create a template stub from natural language description |
delete_template |
Delete a custom template |
interpolate_profiles |
Blend two profiles by linear interpolation |
Preview & Visualization (7)
| Tool | Description |
|---|---|
preview_raw |
Quick preview JPEG with optional inline image return |
preview_before_after |
Side-by-side neutral vs. profile comparison |
preview_exposure_bracket |
Multiple EV stops rendered for exposure comparison |
preview_white_balance |
Multiple WB presets with Kelvin values |
batch_preview |
Thumbnails for multiple RAW files |
preview_luminance_mask |
Grayscale mask showing local adjustment coverage |
preview_with_adjustments |
Preview with all Locallab spots active |
Processing & Export (4)
| Tool | Description |
|---|---|
process_raw |
Process a RAW file to JPEG/TIFF/PNG with inline thumbnail |
apply_template |
Apply a template to process a RAW file with optional device preset |
batch_process |
Process multiple RAW files with the same profile |
export_multi_device |
Export one RAW optimized for multiple devices in one call |
Crop & Device (3)
| Tool | Description |
|---|---|
adjust_crop_position |
Reposition crop (left/center/right, top/center/bottom, or pixel offsets) |
add_device_preset_tool |
Create a custom device preset |
delete_device_preset |
Delete a custom device preset |
Local Adjustments (5)
| Tool | Description |
|---|---|
add_luminance_adjustment |
Add luminance-based local adjustment (shadows/midtones/highlights/custom) |
list_local_adjustments |
List all Locallab spots in a profile |
adjust_local_spot |
Modify an existing Locallab spot |
remove_local_adjustment |
Remove a Locallab spot |
apply_local_preset |
Apply a predefined local adjustment preset with intensity scaling |
Lens Correction (2)
| Tool | Description |
|---|---|
apply_lens_correction |
Apply Lensfun auto-detect or Adobe LCP correction to a PP3 profile |
check_lens_support |
Query the Lensfun database for distortion/vignetting/TCA coverage |
Film Simulation / LUT Support (4)
| Tool | Description |
|---|---|
list_luts |
Scan RT_HALDCLUT_DIR for HaldCLUT film simulation LUTs, grouped by category |
apply_lut |
Write a HaldCLUT film simulation into a PP3 profile with configurable strength |
preview_lut |
Render an inline preview of a RAW file with a film simulation applied |
preview_lut_comparison |
Render 2–5 LUT previews side-by-side for quick comparison |
Profile Inheritance (3)
| Tool | Description |
|---|---|
create_profile_variant |
Derive a child PP3 from a parent template with section-level overrides |
list_profile_variants |
List all variants with override summaries, optionally filtered by parent |
update_base_profile |
Modify a base template and propagate changes to all child variants |
Metadata Privacy (3)
| Tool | Description |
|---|---|
inspect_metadata |
Classify JPEG/TIFF EXIF into sensitive/technical/rights buckets with privacy recommendations |
strip_metadata |
Losslessly remove GPS, serial numbers, software, and owner tags from a JPEG |
set_metadata |
Write copyright, artist, description, and keywords into a JPEG |
Built-in Resources
PP3 Templates (5)
| Template | Description |
|---|---|
neutral |
Minimal processing, camera white balance, basic sharpening |
warm_portrait |
Warm tones (5800K), gentle contrast, skin-friendly saturation |
moody_cinematic |
Cool tones (5200K), lifted blacks, reduced saturation, film look |
vivid_pet |
Warm tones (5600K), boosted saturation and vibrance, strong sharpening |
bw_classic |
Black & white via channel mixer, high contrast, strong sharpening |
Device Presets (19)
- Mobile (7): Samsung Galaxy S26 Ultra, Galaxy S25 Ultra, iPhone 16 Pro Max, iPhone 16, Google Pixel 9 Pro, generic 9:16, generic 9:19.5
- Desktop (5): 4K UHD (3840x2160), WQHD (2560x1440), Full HD (1920x1080), Ultrawide 21:9 (3440x1440), Dual 4K 32:9 (7680x2160)
- Photo Formats (7): 3:2 (35mm), 4:3, 16:9, 1:1 (square), 5:4, 4:5 (Instagram portrait), 2:3 (portrait 35mm)
Custom presets can be added via add_device_preset_tool and persist across sessions.
Local Adjustment Presets (7)
| Preset | Description |
|---|---|
shadow_recovery |
Brighten shadows without affecting highlights |
highlight_protection |
Compress highlights to recover detail |
split_tone_warm_cool |
Warm shadows, cool highlights |
midtone_contrast |
Add contrast to midtones only |
shadow_desaturation |
Desaturate shadow areas for a clean look |
amoled_optimize |
High contrast and deep blacks for AMOLED displays |
hdr_natural |
Natural HDR look with lifted shadows and compressed highlights |
Presets accept an intensity parameter (50 = default, 25 = half, 100 = double).
Environment Variables
| Variable | Default | Description |
|---|---|---|
RT_CLI_PATH |
Auto-detect | Path to rawtherapee-cli binary |
RT_OUTPUT_DIR |
~/Pictures/rawtherapee-mcp-output |
Default output directory |
RT_PREVIEW_DIR |
OS temp dir | Preview image directory |
RT_CUSTOM_TEMPLATES_DIR |
./custom_templates |
Custom PP3 templates directory |
RT_PREVIEW_MAX_WIDTH |
1200 |
Max preview width in pixels |
RT_JPEG_QUALITY |
95 |
Default JPEG quality (1-100) |
RT_LOG_LEVEL |
WARNING |
Logging level (DEBUG, INFO, WARNING, ERROR) |
See .env.example for a documented configuration template.
Known Limitations
- RawTherapee CLI must be installed separately — it is not a Python package and cannot be installed via pip
- stdio transport only — designed for local desktop use, no HTTP/SSE server mode
- RT 5.12 crop+resize bug — RawTherapee 5.12 silently ignores Crop when Resize is also enabled in the same PP3 profile; the server applies a crop-only workaround automatically
- 1MB MCP tool response limit — preview resolution is managed to stay within Claude Desktop's response size limit
- No pixel-level editing — no retouching, object removal, frequency separation, or content-aware fill
- No multi-image compositing — no HDR merge, panorama stitching, or focus stacking
- Luminance masks are tonal only — no spatial or subject-aware masking
- Integration tests require a RawTherapee installation and are skipped by default
Troubleshooting
RawTherapee not found:
- Run
check_rt_statusto see detection details - Set
RT_CLI_PATHto the full path ofrawtherapee-cli:- Windows:
C:\Program Files\RawTherapee\5.12\rawtherapee-cli.exe - macOS:
/Applications/RawTherapee.app/Contents/MacOS/rawtherapee-cli - Linux:
/usr/bin/rawtherapee-cli
- Windows:
Paths with spaces on Windows:
- Always use the full path in
RT_CLI_PATH, including quotes if your shell requires them - Environment variables in MCP client config JSON do not need extra escaping
Preview images too large:
- Reduce
RT_PREVIEW_MAX_WIDTH(default: 1200) - Preview tools auto-thumbnail to stay within the 1MB response limit
Server not responding:
- Verify the entry in your MCP client config is correct
- Restart your MCP client after config changes
- Set
RT_LOG_LEVEL=DEBUGto see detailed logs on stderr
New version installed but old tools still showing in Claude: See docs/TROUBLESHOOTING.md - Stale tool list after update. Short version: fully quit Claude Desktop (it runs as a background process), then relaunch.
Docker
docker build -t rawtherapee-mcp-server .
docker run -i --rm \
-v /path/to/photos:/photos \
-v /path/to/output:/output \
-e RT_OUTPUT_DIR=/output \
rawtherapee-mcp-server
Development
See docs/DEVELOPMENT.md for full setup instructions.
git clone https://github.com/lucamarien/rawtherapee-mcp-server
cd rawtherapee-mcp-server
pip install -e ".[dev]"
make validate # lint + format + typecheck + test + security + audit
Testing
# Unit tests (mocked, no RawTherapee required)
pytest -v
# Integration tests (requires RawTherapee installation)
pytest -m integration -v
MCP Inspector
npx @modelcontextprotocol/inspector uv run rawtherapee-mcp
Contributing
See CONTRIBUTING.md for development guidelines, coding standards, and PR checklist.
License
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.