agent-wiki-documents

Document ingestion MCP server for wiki workflows.

Current coordinated release: 0.14.5.

It converts PDF, Office, text and image files to Markdown. OCR is used for images and for PDFs where direct text extraction returns no useful text.

Supported Inputs

• PDF: MarkItDown text extraction first, LLM OCR for scanned or visually

significant pages.

• Office: Word, PowerPoint and Excel files are converted through MarkItDown.
• Text: txt, md, csv, json, xml, yaml, html, rtf.
• Images: png, jpg, jpeg, tif, tiff, bmp, webp through LLM OCR.

Diagram-like images can include a Mermaid flowchart generated by the same vision LLM pass.

Run Locally

Via llm-wiki-manager (recommended — starts all external agents together):

Register the endpoint in mcp.endpoints.json:

Mounted input files go in:

Generated Markdown files are written to:

Configuration

docker-compose.yml mounts ${WORKSPACES_ROOT} at /workspaces. When documents_convert_to_markdown receives workspace, converted Markdown is ready for wiki ingest in that workspace without a copy step.

MCP Authentication

When running via wiki-workspace agents up, set DOCUMENTS_MCP_AUTH_TOKEN in the manager's .env; it is mapped to MCP_AUTH_TOKEN inside the container.

If MCP_AUTH_TOKEN is empty, the server accepts unauthenticated requests. Use that only for local debugging.