If you’ve been experimenting with AI assistants lately, you’ve probably hit the same wall most people hit: you want to point the AI at a folder of PDFs, a contract, a report, or some internal documents, and ask it questions — but you’re not comfortable uploading those files to a commercial AI service. Maybe they’re client records, financial data, or anything else that belongs on your own machines and nowhere else.
The good news is that this is completely solvable. And in this post, I’ll walk you through exactly how to do it — using free, open-source tools that run entirely on your own computer, with no data ever leaving your network.
This guide is written for developers who are curious but not necessarily experts, and for business owners who want to understand what’s actually possible with private, local AI. By the end, you’ll have a working setup where an AI can read any document on your PC — including scanned PDFs — and answer questions about it, all completely offline.
Why “local AI” is such a big deal for businesses
When you paste a document into ChatGPT or Claude, that text travels to a server somewhere, gets processed, and a response comes back. For casual use, that’s completely fine. But for a law firm handling client communications, an accountant managing financial reports, a healthcare provider dealing with patient records, or any business with confidential internal data — that’s a very different conversation.
Commercial AI services are governed by their own terms of service and data handling policies. Even when a provider promises not to train on your data, you’re still trusting a third party with information that may be subject to GDPR, NDA clauses, industry regulations, or just plain common sense about competitive sensitivity.
Local AI flips the model entirely: the AI runs on your hardware, your documents stay on your drives, and nothing leaves your premises. The intelligence comes to the data, not the other way around.
This isn’t a theoretical future capability. It works today, on ordinary business hardware, and it’s more capable than most people expect.
The tools we’ll use
Here’s the short version of what we’re going to install:
LM Studio — a free desktop app that lets you download and run AI models locally, with a friendly interface and a built-in MCP server for connecting tools.
MCP (Model Context Protocol) — a standard that lets AI assistants connect to external tools and data sources, like your file system or a PDF reader. Think of it as a plugin system for AI.
@modelcontextprotocol/server-filesystem — an official MCP server that gives the AI read access to folders and drives you specify.
@sylphx/pdf-reader-mcp — a lightweight MCP server that extracts text from regular, digitally-created PDFs.
pdf-reader-mcp (Python) — a smarter PDF tool that also handles scanned documents by running OCR using Tesseract, which is a free open-source text recognition engine.
Tesseract OCR — the engine behind the OCR step. It’s been around for decades, is actively maintained, and runs entirely on your machine.
You don’t need a powerful GPU. LM Studio works on most modern laptops and desktops, and for document question-answering, even a mid-range CPU configuration performs well enough to be genuinely useful.
1. Install LM Studio
Head to lmstudio.ai and download the installer for Windows. The installation is straightforward — just run the .exe and follow the prompts.
Once installed, open LM Studio and go to the Discover tab to download a model. If you’re not sure which one to start with, Qwen3.5-9b, Llama 3.1 8B or Mistral 7B Instruct are solid choices that run well on most machines without a dedicated GPU, and for more capable hardware (e.g.at least 64GB of RAM and an NVIDIA 3080) you could even go for Qwen3.5 35B A3B which has reasoning, vision and tools capabilities and can outperform previous models 6x its size.
LM Studio will show you a compatibility indicator so you can see what works for your hardware before you download anything.
Tip: If you’re primarily doing document Q&A rather than creative writing, models labelled “Instruct” tend to be more focused and useful for that task.
After downloading a model, click Load to start it. You’ll see memory usage appear in the status bar at the bottom — that’s your AI running locally, no internet required.
2. Install Tesseract (to read scanned documents)
If you only deal with digital PDFs — documents that were created in Word, exported from software, or are otherwise text-based — you can skip this step. But if you have scanned contracts, archived invoices, photographed receipts, or any document that’s essentially an image of text, you’ll need Tesseract.
Download the Windows installer from the UB-Mannheim project on GitHub — they maintain the official Windows build. Run the installer, and note the path where it installs (usually C:\Program Files\Tesseract-OCR\).
To confirm it worked, open a terminal (press Win + R, type cmd, hit Enter) and run:
tesseract --versionIf you see version info, you’re good to go.
3. Set up full filesystem access via MCP
This is where things get interesting. LM Studio supports MCP, which means you can give it tools to interact with the outside world — including reading files directly from your drives.
The filesystem MCP server from the official Model Context Protocol project handles this. You configure it by telling it which root paths it’s allowed to access. To give your AI access to everything on your PC, just point it at your drive roots.
In LM Studio, navigate to the Developer section, then open your mcp.json configuration file (it lives at ~/.lmstudio/mcp.json on Windows, which translates to C:\Users\YourName\.lmstudio\mcp.json). Add this block inside the mcpServers section:
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"C:\\",
"D:\\",
"E:\\"
]
}Adjust the drive letters to match what you actually have. This gives the AI read access to your entire filesystem — it can navigate directories, open files, search for content, and read anything you ask it to.
Keep in mind: This MCP server is read-only by default, so the AI can read your files but can’t modify or delete them. That’s an important safety boundary that’s baked into the tool.
4. Add a PDF reader with OCR
The filesystem MCP is great for plain text files, but PDFs are a different beast. They have their own internal format, embedded fonts, and scanned pages are basically just images. We need a dedicated PDF tool.
We’ll install two complementary servers — one fast and lightweight for regular PDFs, and one smarter one for scanned documents.
For standard digital PDFs
Add this to your mcp.json:
"pdf-reader": {
"command": "npx",
"args": ["-y", "@sylphx/pdf-reader-mcp"]
}This one is zero-configuration, pulls in no external dependencies, and handles most everyday PDFs very well.
For scanned or image-based PDFs
First, make sure you have Python installed (download from python.org if not — get Python 3.10 or higher). Then open a terminal and run:
pip install "pdf-reader-mcp[ocr] @ git+https://github.com/damateosg/pdf-reader-mcp.git"This installs a smarter PDF server that uses PyMuPDF to render pages and Tesseract to read the text from scanned images. Then add it to your mcp.json:
"pdf-reader-ocr": {
"command": "C:\\Users\\YourName\\AppData\\Local\\Programs\\Python\\Python313\\Scripts\\pdf-reader-mcp.exe",
"env": {
"TESSERACT_CMD": "C:\\Program Files\\Tesseract-OCR\\tesseract.exe"
}
}Replace YourName with your actual Windows username, and adjust the Python path if your installation is in a different location.
where pdf-reader-mcpin a terminal will tell you the exact path.
No PDF OCR mpc: the personal AI cannot detect content in the PDF.
The smart thing about this server is the read_pdf_smart tool — it automatically detects whether a PDF has embedded text or is image-based, and applies the right extraction method. You don’t need to think about which one to use.
5. Put it all together
Restart LM Studio after saving your mcp.json. When you start a new chat, you should see a tools icon or MCP status indicator confirming the servers are connected.
Now try asking your local AI something like:
- “Read the PDF at C:\Documents\contract_2025.pdf and summarise the payment terms.”
- “Search E:\Invoices\ for any invoice mentioning the company Acme Corp and list the amounts.”
- “Open the scanned document at D:\Archives\old_agreement_scan.pdf and tell me the key dates mentioned.”
- “Read all the PDFs in the folder E:\Payments\ and build me a single calendar file with deadlines and amounts.“
The AI will use the filesystem MCP to navigate to the file, hand it to the appropriate PDF reader, and give you a grounded, accurate answer based on the actual contents of that document — not a hallucinated guess.
With the PDF OCR mcp: now the personal AI can read the contents of a scanned PDF seamlessly.
Tip: The OCR reader has a read_pdf_smart tool that works as a catch-all. If you’re not sure whether a PDF is scanned or digital, just use that one and let it figure it out.
The business case for document privacy
What we’ve just built is the foundation of something genuinely valuable for any business that deals with sensitive paperwork — which is most of them.
Think about what your team does with documents every day: reviewing contracts, comparing supplier quotes, extracting data from invoices, searching through compliance records, summarising meeting notes. These are time-consuming, repetitive tasks that AI is extremely good at. But they involve data you should not be sending to a third-party server.
With a locally-running AI connected to your filesystem and document tools, you get the productivity gains without the privacy risk. Your documents stay on your machines. Your data stays in your network. You’re not subject to another company’s data retention policies or potential breach risks.
This matters even more when you consider regulated industries. If your business handles medical records, legal correspondence, financial data, or anything covered by GDPR, HIPAA, or sector-specific compliance frameworks, the question isn’t just “is this convenient” — it’s “is this compliant.”
The setup described in this article is what a real private AI deployment looks like at its most basic level. The same principles — local models, controlled data access, no external API calls — scale up to full enterprise deployments with multiple users, role-based access, audit logging, and integration with your existing business systems.
Taking it further with custom AI systems for Business
What we’ve walked through here is a working single-user setup. It’s genuinely useful as-is, and it gives you a real sense of what local AI can do. But if you’re thinking about deploying this kind of capability across a team, or building it into a business workflow, there’s quite a bit more to consider:
- Model selection matters a lot for document work. Some models are much better at structured extraction and summarisation than others, and the right choice depends on your document types and languages.
- RAG (Retrieval-Augmented Generation) lets you build a searchable knowledge base from your document library, so the AI can answer questions across hundreds of documents at once, not just one at a time.
- Multi-user deployments need proper access controls, so different people see only the documents they’re supposed to see.
- Integration with your existing tools — your CRM, ERP, file shares, or case management system — is where the real productivity gains come from.
These are exactly the kinds of problems we solve at Shambix. If you’re a developer, this guide gives you enough to start exploring on your own. If you’re a business owner or CTO thinking about what a production-grade private AI deployment would look like for your organisation, we’d love to talk through your specific situation.