Search Your Company File Server with AI: PDF, Excel and Outlook Files

The problem with "chat with your documents"

Most companies keep their knowledge in folders: price lists from suppliers, technical data sheets, contracts, offers, e-mails saved as .msg, Excel exports. Uploading that into a chat does not scale, and the usual "knowledge base" products cut every document into fragments and search them by meaning. That works for prose. It works badly for what people actually look for in a file server: an article number, a customer name, a norm, a price row, the newest version of a document.

AnythingMCP takes a different route. It indexes the full text of every page, keeps the folder structure, the file names and dates, and gives your AI assistant tools to search the way a careful colleague would: find the right file, open the right page, quote the row, name the source.

How it works

1. Connect a folder. Self-hosted: a share or volume mounted into the AnythingMCP container. Cloud: an SFTP, FTPS or WebDAV server.
2. AnythingMCP indexes it. Every supported file is converted to text: PDFs page by page, Excel row by row with the column names, Outlook messages with their attachments. New and changed files are picked up on every sync; unchanged files are skipped.
3. Your AI assistant searches it. Through MCP tools: search by content, find by name, list a folder, read pages, exact search for codes, and the file itself as a download link.

Self-hosted: mount the share

This is the setup for a Windows file server, a NAS or any SMB/CIFS or NFS share in your network.

1. Mount the share read-only on the Docker host, for example /mnt/knowledge (with cifs-utils: //fileserver/group$/Knowledge /mnt/knowledge cifs ro,credentials=/root/.smb-knowledge,iocharset=utf8 0 0 in /etc/fstab).

2. Bind it read-only into the app and files-worker containers, and allow it as a file root with FILE_SOURCE_ROOTS=/mnt/knowledge in .env:

   services:
     app:
       volumes:
         - /mnt/knowledge:/mnt/knowledge:ro
     files-worker:
       volumes:
         - /mnt/knowledge:/mnt/knowledge:ro
   

3. Turn on File Search: set COMPOSE_PROFILES=files, FILES_WORKER_URL=http://files-worker:8891 and two random tokens, FILES_WORKER_TOKEN and FILE_EXTRACTOR_TOKEN (openssl rand -hex 24), then docker compose up -d. The File Search docs list every setting.

4. In AnythingMCP, open Connectors → New → File Search, choose Network folder and follow the wizard: pick /mnt/knowledge, click Scan folder to see every subfolder with its number of files (archive folders come unticked), check that the selection fits your plan, review the description AnythingMCP drafts from the scan, and click Activate.

The files never leave your network. Only what a tool returns for a question (a list of matches, the text of a page) travels to the AI provider you use.

AnythingMCP Cloud: SFTP, FTPS or WebDAV

On the Cloud, AnythingMCP reaches your files over the internet. Most NAS systems (Synology, QNAP, Asustor) and file servers offer SFTP or WebDAV out of the box; Nextcloud and ownCloud speak WebDAV. Set up a read-only account for the folder and follow the guide for your source: SFTP, FTP / FTPS, WebDAV, Nextcloud, ownCloud, Synology NAS, QNAP NAS, Fritz!Box storage, Hetzner Storage Box or an S3 bucket.

It works in every AI client that supports remote MCP servers: Claude, ChatGPT, Microsoft Copilot, Gemini, Cursor and Meta Muse.

What makes the answers reliable

• Page-level results. A search hit names the file, the page and the file's date. The assistant reads that page instead of guessing from a fragment.
• Exact search. Article numbers, order numbers and norms (DIN 18101, NT-601, 4711-200) are found literally, including inside Excel rows.
• Newest first when it matters. Results carry the modification date, and the assistant is told to prefer the newest of several versions and to say how old a document is.
• Old versions stay out. Exclude folders such as Archive or Old and they are never indexed.
• Unreadable files are reported, not hidden. Password-protected or damaged files stay findable by name, and the Files tab lists why they could not be read.

Permissions

• Per tool: MCP roles decide who gets the file tools at all.
• Per connector: connect two folders as two connectors (for example "Knowledge base" for everyone, "Supplier conditions" for purchasing) and assign them to different roles.
• Per folder (Enterprise): allow or deny folders inside one connector for each role.
• Audit: every search, read and download link is logged with the user, as for every AnythingMCP tool.

Which files are supported

Text is extracted from over 100 file types: PDF page by page; Word, Excel (row by row with the column names) and PowerPoint, including macro and template variants and .xlsb; OpenDocument; Outlook .msg, .eml and .mbox with the text of their attachments; .rtf, .epub, .xps; Apple Pages, Numbers and Keynote through their PDF preview; .svg and the text in .dxf CAD drawings; CSV, Markdown, HTML, XML, JSON, YAML, calendars, contacts and source code. Every other file is still found by name and folder. For what text cannot capture (a drawing's title block, a stamp, a photo), the AI can look at a PDF page or an image as a picture. Scans without a text layer are read with OCR on Enterprise.

Prompts to try

• "Which supplier documents mention fire protection class T30, and which is the most recent?"
• "Find the e-mail from Müller GmbH about the delivery delay in March and summarise what was agreed."
• "Show me every Excel row in the folder Conditions that contains the discount group 12."
• "What does our internal procedure say about returns older than 30 days? Quote the paragraph and the file."

Plans

On the Cloud and self-hosted, Starter includes one File Search connector with up to 500 files, Team three connectors with 2,000 files, both synced daily. The Cloud trial includes one connector with up to 500 files and files up to 10 MB, synced daily, without File Share links. Enterprise removes the limits, syncs down to every 15 minutes (5 minutes self-hosted) and adds OCR, folder permissions per role and workspace policies. The free self-hosted Community edition can try File Search during the 30-day trial, which has the Team limits; after that it needs a licence.

If the folder grows past your plan, the sync pauses and the connector page says so, with the number of files against the limit; the AI keeps searching what is already indexed until you narrow the selection or upgrade. Compare plans.

FAQ

Do the files go to OpenAI or Anthropic? No. AnythingMCP keeps the index on your server (or in your Cloud workspace). The AI client only receives what a tool returns for a question: a list of matches, the text of a page, a link.

How big can the share be? Self-hosted, the limits are your hardware and your plan's number of files. Indexing runs in its own background service and does not slow down the tools.

What about Google Drive, OneDrive or SharePoint? Use their own connectors in the catalog; they search through the provider's API.

Related

• Find prices in supplier price lists with AI
• Search inside Nextcloud and WebDAV files with AI
• Share files from an AI chat as download links
• Self-hosted network folder in Claude, ChatGPT, Microsoft Copilot, Gemini, Cursor, Muse
• File sources one by one, in Claude: SFTP, FTP / FTPS, WebDAV, S3 bucket, Nextcloud, ownCloud, Synology NAS, QNAP NAS, Fritz!Box storage, Hetzner Storage Box, network folder (self-hosted)
• Connect an FTP or SFTP server to Claude, ChatGPT, Microsoft Copilot, Gemini, Cursor or Muse