mcp-server-sdlxliff

Chat with your SDL Trados translation files using Claude.

Review translations, find errors, and make corrections through simple natural conversation - no manual XML editing required.

Two Ways to Use

Option	Best For	Claude Access
XLIFF Chat Desktop App	Professional users who want a dedicated app	API key (pay per use)
Claude Desktop Extension	Claude Pro/Team subscribers	Claude subscription

Both options use the same MCP server under the hood - choose based on how you prefer to access Claude.

What You Can Do

Just ask Claude in plain language:

• "Check this file for grammar errors in the Russian translations"
• "Find segments where the translation is much longer than the source"
• "Fix segment 42 - change 'программа' to 'приложение'"
• "Show me all segments that are still in Draft status"
• "Run QA checks with my glossary"
• "Save my changes"

Claude reads your SDLXLIFF files, understands the translation context, and can make corrections while preserving all formatting tags automatically.

XLIFF Chat Desktop App

A standalone macOS app with a native file picker - no command line required.

Installation

1. Install the MCP server globally (not in a virtual environment):

   pip3 install mcp-server-sdlxliff
   

   Note: The desktop app looks for Python at /opt/homebrew/bin/python3 (Apple Silicon) or /usr/local/bin/python3 (Intel Mac). Make sure to install with the same pip3 that corresponds to your Homebrew Python.

2. Download XLIFF Chat_1.1.0_aarch64.dmg from Releases

3. Open the .dmg and drag XLIFF Chat to your Applications folder

4. Launch XLIFF Chat and enter your Anthropic API key

Usage

1. Click File or Folder button to select SDLXLIFF files
2. The MCP server connects automatically
3. Start chatting - ask Claude to review, check, or edit your translations

Requirements

• macOS 10.15 or later
• Python 3.10+ with mcp-server-sdlxliff installed globally
• Anthropic API key

Claude Desktop Extension

Use with your Claude Pro or Team subscription through Claude Cowork.

Installation

Option A: Desktop Extension (Recommended)

1. Download mcp-server-sdlxliff-1.1.0.mcpb from Releases
2. Open Claude Desktop → Settings → Extensions
3. Click "Install Extension" and select the downloaded .mcpb file
4. The extension installs automatically (Python and dependencies are managed for you)

Option B: Manual Installation

pip install mcp-server-sdlxliff

Then add to your claude_desktop_config.json:

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

{
  "mcpServers": {
    "sdlxliff": {
      "command": "mcp-server-sdlxliff"
    }
  }
}

Usage with Claude Cowork

1. Open Claude Cowork
2. Click "+ Add folder" and select a folder containing your SDLXLIFF files
3. The MCP tools will be available automatically
4. Ask Claude to review or edit translations

Compatibility

Product	Works?	Notes
Claude Cowork	Yes	Add a folder with SDLXLIFF files and start chatting
Claude Desktop (chat)	No	Sandbox limitation prevents file access
Claude.ai (web)	No	No local filesystem access

Features

• Natural conversation - Ask questions and request changes in plain language
• Error detection - Find grammar, spelling, consistency, and terminology issues
• QA checks - Automated quality assurance with glossary support
• Safe corrections - Edit translations while preserving all formatting tags
• Batch review - Process large files with automatic pagination
• Change tracking - Modified segments are marked as RejectedTranslation for easy review in Trados

Human in the Loop

Claude's corrections are suggestions, not final changes. The workflow keeps you in control:

1. Claude marks all modified segments as RejectedTranslation
2. Open the file in SDL Trados Studio
3. Filter by status → RejectedTranslation to see only Claude's changes
4. Review each suggestion and Confirm or Edit as needed
5. Your approved changes become Translated or ApprovedTranslation

This ensures human oversight - you always have the final say on what goes into the translation.

Available Tools

read_sdlxliff

Extract translation segments from an SDLXLIFF file.

Parameters:

• file_path (string, required): Path to the SDLXLIFF file
• offset (integer, optional): Starting segment index for pagination (default: 0)
• limit (integer, optional): Maximum segments to return (default: all). Use 50 for large files.
• include_tags (boolean, optional): Include tagged text fields for segments with formatting tags (default: false)

Returns: JSON object with pagination metadata and segments:

{
  "total_segments": 184,
  "offset": 0,
  "count": 50,
  "has_more": true,
  "segments": [...]
}

Each segment contains:

• segment_id: Unique segment identifier (mrk mid)
• trans_unit_id: Parent trans-unit ID
• source: Source text (clean, without tags)
• target: Target text (clean, without tags)
• has_tags: Whether segment contains inline formatting tags
• source_tagged: Source with tag placeholders (only if has_tags=true and include_tags=true)
• target_tagged: Target with tag placeholders (only if has_tags=true and include_tags=true)
• status: SDL confirmation level (e.g., Translated, RejectedTranslation)
• locked: Whether segment is locked
• percent: TM match percentage, e.g., 100 for exact match (only present when available)
• origin: Match origin, e.g., tm, document-match, auto-propagated (only present when available)
• text_match: Context Match indicator - value SourceAndTarget means this is a CM (only present for context matches)
• repetitions: Number of times this source text appears in the file (only present when > 1)

get_sdlxliff_segment

Get a specific segment by its ID.

Parameters:

• file_path (string, required): Path to the SDLXLIFF file
• segment_id (string, required): The segment ID (mrk mid) to retrieve

update_sdlxliff_segment

Update a segment's target text. Automatically sets status to RejectedTranslation.

Parameters:

• file_path (string, required): Path to the SDLXLIFF file
• segment_id (string, required): The segment ID (mrk mid) to update
• target_text (string, required): New target text. For segments with tags, include placeholders.
• preserve_tags (boolean, optional): Validate and restore tags from placeholders (default: true)

Note: Changes are kept in memory until save_sdlxliff is called. For segments with formatting tags (has_tags=true), you must include tag placeholders in the target text. See Tag Handling below.

validate_sdlxliff_segment

Pre-validate proposed changes to a segment before updating.

Parameters:

• file_path (string, required): Path to the SDLXLIFF file
• segment_id (string, required): The segment ID to validate against
• target_text (string, required): Proposed target text with tag placeholders

Returns: Validation result with errors, warnings, and missing/extra tag information.

save_sdlxliff

Save changes to the SDLXLIFF file.

Parameters:

• file_path (string, required): Path to the SDLXLIFF file
• output_path (string, optional): Alternative output path (default: overwrites original)

get_sdlxliff_statistics

Get statistics about the translation file.

Parameters:

• file_path (string, required): Path to the SDLXLIFF file

Returns: JSON object with:

• total_segments: Total number of segments
• status_counts: Count of segments by SDL confirmation level
• locked_count: Number of locked segments

qa_check_sdlxliff

Run quality assurance checks on the translation file.

Parameters:

• file_path (string, required): Path to the SDLXLIFF file
• segment_ids (array of strings, optional): Specific segment IDs to check. If omitted, checks all segments.
• checks (array of strings, optional): Specific checks to run. If omitted, runs default checks (all except spelling).
• glossary_path (string, optional): Path to glossary file for terminology check. If omitted, auto-discovers glossary.tsv, glossary.txt, terminology.tsv, or terminology.txt in the same directory as the SDLXLIFF file.
• dictionary_path (string, optional): Path to custom dictionary for spelling check. If omitted, auto-discovers dictionary.txt, custom_words.txt, or spelling.txt in the same directory as the SDLXLIFF file.

Available checks:

Check	Description	Default
trailing_punctuation	Source ends with .!?:; but target doesn't (or vice versa)	Yes
numbers	Numbers in source don't match numbers in target	Yes
double_spaces	Target contains consecutive spaces	Yes
whitespace	Leading/trailing whitespace mismatch between source and target	Yes
brackets	Different count of ()[]{} between source and target	Yes
inconsistent_repetitions	Segments with same source text have different translations	Yes
terminology	Glossary terms from source must appear in target (requires glossary file)	Yes
spelling	Spellcheck target text using target language from file metadata	No (opt-in)

Spelling check:

• Must be explicitly requested: checks: ["spelling"] or checks: ["spelling", "numbers", ...]
• Uses target language from SDLXLIFF metadata (e.g., de-DE, ru-RU)
• Supported languages:
  • Russian, Ukrainian, English: Yandex.Speller (proper morphological dictionary)
  • German, Spanish, French, Italian, Portuguese, Dutch: pyspellchecker
• Unsupported languages are silently skipped (no false positives for rare languages)

Glossary file format:

# Comment lines start with #
# Format: source_term<TAB>target_term
Galaxy	Galaxy
Settings	Настройки
Smart Switch	Smart Switch

Single terms (without tab) mean the term must appear unchanged in the target.

Custom dictionary file format (for spelling):

# Comment lines start with #
# One word per line (case-insensitive)
SDLXLIFF
Trados
MCP
ProductName

Returns: JSON object with:

{
  "total_segments": 184,
  "segments_checked": 184,
  "segments_with_issues": 12,
  "glossary_used": "/path/to/glossary.tsv",
  "glossary_terms_count": 5,
  "target_language": "de-DE",
  "dictionary_used": "/path/to/dictionary.txt",
  "custom_words_count": 10,
  "issues": [
    {
      "segment_id": "42",
      "check": "spelling",
      "severity": "warning",
      "message": "Possible misspelling: 'korrektr' (suggestions: korrektor, korrekter, korrekt)",
      "source_excerpt": "",
      "target_excerpt": "Das ist ein korrektr Satz..."
    }
  ],
  "summary": {
    "trailing_punctuation": 5,
    "numbers": 3,
    "spelling": 2
  }
}

If spelling is requested but the language is unsupported:

{
  "spelling_skipped": "Language 'ja-JP' not supported for spelling check"
}

SDLXLIFF Format

SDLXLIFF is SDL's extension of the XLIFF 1.2 standard, used by SDL Trados Studio. Key characteristics:

• Each <mrk mtype="seg"> element is a separate translatable segment
• Segment IDs are the mid attribute values (globally unique numbers)
• Status is stored in <sdl:seg conf="..."> (not the XLIFF state attribute)
• Valid SDL confirmation levels: Draft, Translated, RejectedTranslation, ApprovedTranslation, RejectedSignOff, ApprovedSignOff

Tag Handling

SDLXLIFF files often contain inline formatting tags (<g>, <x>, <bpt>, <ept>, etc.) that control bold, italic, colors, line breaks, and other formatting in the final document. These tags must be preserved during translation updates or the formatting will be lost.

How It Works

The MCP server converts XML tags to readable placeholders:

XML Tag	Placeholder	Description
<g id="5">text</g>	{5}text{/5}	Paired formatting tag
<x id="5"/>	{x:5}	Self-closing tag (e.g., line break)
<bpt id="5">	{5}	Begin paired tag
<ept id="5">	{/5}	End paired tag

Example

Original XML:

<mrk mid="3"><g id="5">Acme</g><g id="6">&amp;</g><g id="7"> Events</g></mrk>

Extracted as:

{
  "source": "Acme& Events",
  "source_tagged": "{5}Acme{/5}{6}&{/6}{7} Events{/7}",
  "has_tags": true
}

To update this segment, you must include all tags:

{5}Acme{/5}{6}&{/6}{7} Мероприятия{/7}

Validation Rules

When updating segments with tags:

1. All original tags must be present - Missing tags will cause the update to be rejected
2. No extra tags allowed - Only tags from the original segment can be used
3. Tags must be properly paired - Opening {5} must have matching {/5}
4. Tag order can change - Word order differences between languages are allowed (with a warning)

Workflow for Tagged Segments

1. Read segments with read_sdlxliff - check has_tags field
2. For segments with has_tags=true, use get_sdlxliff_segment to get the source_tagged/target_tagged fields
3. Include all placeholders when calling update_sdlxliff_segment
4. If validation fails, the error message shows which tags are missing

Claude Desktop Chat Limitation

Claude Desktop Chat runs Claude in a gVisor sandboxed container for security. When you attach files via "+ Add files", they are uploaded to /mnt/user-data/uploads/ inside this container.

The problem: MCP servers run on your host machine, not inside the container. They cannot access the sandboxed filesystem where attached files are stored.

┌─────────────────────────────────────────┐
│          gVisor Container               │
│  ┌─────────────────────────────────┐    │
│  │  /mnt/user-data/uploads/        │    │
│  │    └── your_file.sdlxliff       │ ←── File uploaded here
│  └─────────────────────────────────┘    │
│              Claude                      │
└─────────────────────────────────────────┘
           ╳ No access ╳
┌─────────────────────────────────────────┐
│           Host Machine                   │
│  ┌─────────────────────────────────┐    │
│  │      MCP Server (sdlxliff)      │ ←── Cannot read container filesystem
│  └─────────────────────────────────┘    │
└─────────────────────────────────────────┘

Workaround: Use Claude Cowork or XLIFF Chat instead, which have direct access to your local files.

Development

# Clone the repository
git clone https://github.com/EugeneAnt/mcp-server-sdlxliff.git
cd mcp-server-sdlxliff

# Install with dev dependencies
uv pip install -e ".[dev]"

# Run tests
pytest

Building the Desktop Extension

To create a .mcpb bundle for distribution:

# Install the MCPB CLI
npm install -g @anthropic-ai/mcpb

# Validate the manifest
mcpb validate manifest.json

# Create the bundle
mcpb pack .

This creates mcp-server-sdlxliff-X.X.X.mcpb ready for installation in Claude Desktop.

Building the Desktop App

cd desktop

# Install dependencies
bun install

# Build for macOS
bun run tauri build

This creates XLIFF Chat.dmg in desktop/src-tauri/target/release/bundle/dmg/.

Acknowledgments

• Spelling check for Russian, Ukrainian, and English: Yandex.Speller

License

MIT License - see LICENSE for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.