FLExTools MCP

An MCP server that enables AI assistants to write FLExTools scripts and directly manipulate FieldWorks lexicon data using natural language.

Developed for SIL Global by Matthew Lee in connection with the SIL's AI Integration Advisory Board and the FLExTrans team.

Quick Overview

What it does: FLExTools MCP gives AI assistants (Claude, Copilot, Gemini) the knowledge to write FLExTools modules by providing indexed, searchable documentation of LibLCM and FlexLibs APIs.

** Videos **

Videos

The MCP Connection: Talking to your Dictinary

This podcast, for a linguistic audience, gives an overview of using the FLExTools MCP.

MCPs for FLEx and FLExTools: LangTech AI Software Engineering CoP

This presentation, given to an audience of programmers, discusses the background, architecture, and advantages of an MCP, and introduces the FLExTools MCP.

Three ways to use it:

1. Generate legacy modules (FlexLibs stable)
2. Generate modern modules (FlexLibs 2.0 with ~1,400 functions)
3. Run operations directly on FieldWorks databases using natural language queries

Example: "Delete any sense with 'q' in the gloss" → AI generates, tests, and runs the operation automatically.

⚠️ Warning: Backup your project first - there are no guard-rails.

Why MCP? Why AI?

• What is an MCP Server? See WHY-MCP.md - explains the LibLCM complexity problem and why generic AI assistants fail
• When is AI useful? See WHY-AI.md - learning curve problems and when manual approaches are better

Getting Started

1. Installation

See SETUP.md for detailed installation steps.

Quick summary:

git clone https://github.com/MattGyverLee/FlexToolsMCP.git
cd FlexToolsMCP
pip install -r requirements.txt

# Install FlexLibs 2.0
cd ..
git clone https://github.com/MattGyverLee/flexlibs.git flexlibs2
pip install ./flexlibs2

# Test it works
cd FlexToolsMCP
python -c "from src.server import APIIndex, get_index_dir; i=APIIndex.load(get_index_dir()); print('Loaded', len(i.flexlibs2.get('entities', {})), 'FlexLibs2 entities')"

2. Connect to Your AI Assistant

See SETUP.md for Claude Code, Antigravity, and other tools.

Note: Each AI tool has different MCP configuration syntax. See SETUP.md for your specific tool.

3. Updating to New Versions

See SETUP.md for how to pull new releases.

4. Start Using

See USAGE.md for workflows, tool reference, and examples.

What's Included

MCP Tools (16)

Admin & Config:

• flextools_start - Initialize session, set project and API mode
• flextools_manage_config - Get/set/delete persistent configuration
• flextools_get_session_history - View operation history and undo stack
• flextools_undo_last_operation - Undo the most recent write
• flextools_get_module_template - Get FLExTools module boilerplate

Discovery:

• flextools_search_by_capability - Find APIs by natural language intent
• flextools_get_object_api - Get full API for an object/operations class
• flextools_get_navigation_path - Find traversal between object types
• flextools_find_examples - Get code examples by operation type
• flextools_resolve_property - Check casting requirements for properties

Catalog:

• flextools_list_categories - List semantic domains (lexicon, grammar, etc.)
• flextools_list_entities_in_category - List entities in a domain

Module & Execution:

• flextools_start_module - Interactive wizard for new module
• flextools_get_operation_logs - View logs and pattern recommendations
• flextools_run_module - Execute code with dry-run and write modes

API Coverage

• LibLCM: 2,295 C# entities
• FlexLibs Stable: ~71 methods
• FlexLibs 2.0: ~1,400 methods (99% documented, 82% with examples)

Test-Proven Examples

"Remove 'el ' from the beginning of any Spanish gloss"
"Add an environment named 'pre-y' with the context '/_y'"
"Delete the entry with lexeme ɛʃːɛr"
"List entries with "ː" in the headword"
"Are there any duplicates by gloss (fuzzy match) and POS?"

Key Features

• Discovery-first workflow - the AI assembles modules from indexed building blocks (signatures, navigation skeletons, examples, casting fixes) rather than inventing API calls from training memory. See USAGE.md.
• Automatic index refresh when you update FieldWorks or libraries
• Dry-run mode to test before writing data
• Semantic search with synonym expansion
• Pythonnet casting detection - warns when you need type conversions
• Code examples extracted from real-world usage
• Multiple library versions supported simultaneously

Documentation

Document	Purpose
HISTORY.md	Release notes and version history
SETUP.md	Installation and AI tool configuration
USAGE.md	How to use the MCP, workflows, examples
DEVELOPMENT.md	Project structure, architecture, contributing
docs/WHY-MCP.md	Why FieldWorks needs MCP servers
docs/WHY-AI.md	When AI is useful for FieldWorks work
docs/INNOVATIONS.md	Technical innovations in this MCP
docs/BACKGROUND.md	Project history

Safety & Limitations

Safety

• Always backup before write operations - the MCP defaults to dry-run mode
• Dry run shows what would happen before writing
• Requires explicit user permission for write operations

Limitations

• Cannot control the FLEx GUI (filters, display, etc.)
• Only manipulates data, not UI state
• FlexLibs 2.0 still undergoing extensive testing
• Some Scripture module edge cases recently fixed

Architecture

User Request -> AI Assistant -> MCP Server -> Indexed APIs
                    |
            Generated FLExTools Script or Direct Execution
                    |
            FLExTools (IronPython) or FlexLibs 2.0
                    |
            LibLCM (C# data model)
                    |
            FieldWorks Database

For technical details, see DEVELOPMENT.md.

License

MIT License - See LICENSE file for details

Contributing

Contributions are welcome! Please submit issues and pull requests on GitHub.

For development info, see DEVELOPMENT.md.

Acknowledgments

• The FieldWorks developers (Jason, Ken, Hasso, and team)
• Craig, the developer of FLExTools and FlexLibs
• The SIL AI Implementation Advisory Board
• Ron, Beth and the FLExTrans team
• My mentors Doug, Jeff, and Jenni at SIL LangTech