Privacy-first document processing FastMCP server. Extract, anonymize, and segment documents through FastMCP's modern tool interface.
Please note: As an MCP, privacy of file contents cannot be absolutely guaranteed, but it is a central design consideration. While file contents should be low risk (but non-zero) risk for leakage, file names will, unavoidably and by design, be read and written by the MCP. Plan accordingly. Consider using a local model.
# Install via pip
pip install inkognito
# Or via uvx (no Python setup needed)
uvx inkognito
# Or run directly with FastMCP
fastmcp run inkognitoIf not already present, you need to make sure you add a filesystem MCP.
Add to your claude_desktop_config.json:
{
"mcpServers": {
"inkognito": {
"command": "uvx",
"args": ["inkognito"],
"env": {
// Optional: Add keys when extractors are implemented
// "AZURE_DI_KEY": "your-key-here",
// "LLAMAPARSE_API_KEY": "your-key-here"
}
},
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/you/input-files-or-whatever",
"/Users/you/output-folder-if-you-want-one"
],
"env": {},
"transport": "stdio",
"type": null,
"cwd": null,
"timeout": null,
"description": null,
"icon": null,
"authentication": null
}
}
}In Claude Desktop:
"Extract this PDF to markdown"
"Anonymize all documents in my contracts folder"
"Split this large document into chunks for processing"
"Create individual prompts from this documentation"
- Universal PII detection (50+ types)
- Consistent replacements across all documents
- Reversible with secure vault file
- No configuration needed - smart defaults
- Available Now: Docling (default, with OCR support)
- Planned: Azure DI, LlamaIndex, MinerU (placeholders only)
- Auto-selects best available option
- Falls back to Docling if no cloud options
- Large documents: 10k-30k token chunks
- Prompt generation: Split by headings
- Preserves context and structure
- Markdown-native processing
All tools are exposed through FastMCP's modern interface with automatic progress reporting and error handling.
Replace PII with consistent fake data across multiple files.
anonymize_documents(
directory="/path/to/docs",
output_dir="/secure/output"
)Convert PDF/DOCX to markdown.
extract_document(
file_path="/path/to/document.pdf",
extraction_method="auto" # auto, docling (others coming soon)
)Split large documents for LLM processing.
segment_document(
file_path="/path/to/large.md",
output_dir="/output/segments",
max_tokens=20000
)Create individual prompts from structured content.
split_into_prompts(
file_path="/path/to/guide.md",
output_dir="/output/prompts",
split_level="h2", #configurable, LLM should be able to read the contents of these files safely
)Restore original PII using vault.
restore_documents(
directory="/anonymized/docs",
output_dir="/restored",
vault_path="/secure/vault.json"
)| Extractor | Status | Notes |
|---|---|---|
| Docling | ✅ Fully Implemented | Default extractor with OCR support (OCRMac on macOS, EasyOCR on other platforms) |
| Azure DI | Requires AZURE_DI_KEY environment variable when implemented |
|
| LlamaIndex | Requires LLAMAPARSE_API_KEY environment variable when implemented |
|
| MinerU | Will require magic-pdf library when implemented |
Following FastMCP conventions, all configuration is via environment variables:
# Optional API keys for cloud extractors (when implemented)
export AZURE_DI_KEY="your-key-here"
export LLAMAPARSE_API_KEY="your-key-here"
# Optional OCR languages (comma-separated, default: all available)
export INKOGNITO_OCR_LANGUAGES="en,fr,de"You: "Anonymize all contracts in the merger folder for review"
Claude: "I'll anonymize those contracts for you...
[Processing 23 files...]
✓ Anonymized 23 contracts
✓ Replaced: 145 company names, 89 person names, 67 case numbers
✓ Vault saved to: /output/vault.json
You: "Extract this 300-page research PDF"
Claude: "I'll extract that PDF to markdown...
[Using Docling for extraction...]
✓ Extracted 300 pages
✓ Preserved: tables, figures, citations
✓ Output size: 487,000 tokens
✓ Saved to: research_paper.md
You: "Split this API documentation into individual prompts"
Claude: "I'll split the documentation by endpoints...
[Splitting by H2 headings...]
✓ Created 47 prompt files
✓ Each prompt includes endpoint context
✓ Ready for training or testing
| Extractor | Speed | Requirements | Status |
|---|---|---|---|
| Azure DI | 0.2-1 sec/page | API key | Planned |
| LlamaIndex | 1-2 sec/page | API key | Planned |
| MinerU | 3-7 sec/page | Local, GPU | Planned |
| Docling | 5-10 sec/page | Local, CPU | ✅ Available |
- Local processing: No cloud services required
- No persistence: Nothing saved without explicit paths
- Secure vaults: Encrypted mapping storage
- API key safety: Never logged or transmitted
# Clone the repository
git clone https://github.com/phren0logy/inkognito
cd inkognito
# Run with FastMCP CLI
fastmcp dev
# Or run directly in development
uv run python server.py# Install the server configuration
fastmcp install inkognito
# Test a specific tool
fastmcp test inkognito extract_documentinkognito/
├── pyproject.toml # FastMCP-compatible packaging
├── LICENSE # MIT license
├── README.md # This file
├── server.py # FastMCP server and entry point
├── anonymizer.py # PII detection and anonymization
├── vault.py # Vault management for reversibility
├── segmenter.py # Document segmentation
├── exceptions.py # Custom exceptions
├── extractors/ # PDF extraction backends
│ ├── __init__.py
│ ├── base.py
│ ├── registry.py
│ ├── docling.py # ✅ Implemented
│ ├── azure_di.py # Placeholder
│ ├── llamaindex.py # Placeholder
│ └── mineru.py # Placeholder
└── tests/
MIT License - see LICENSE file for details.