Give AI access to your private documents without forcing them into the cloud. Compress, merge, redact and prepare PDF, Word or Excel in your browser, and connect the same engine to Claude, Cursor or your own agent.
Try it now: PDF to Markdown for AI
21
tools, all inside your browser
0
bytes of your file sent anywhere
0
accounts to create
11 MB
of OCR engine, hosted right here
For AI agents
Your agent reasons; Sırdaş hands it the documents. An MCP server that runs on your machine reads PDFs, scans, Word, Excel, email and photos, and gives the model what it needs, with the page every fact comes from.
claude mcp add sirdas -- npx -y @sirdaspdf/mcp@0.4.3New in version 0.4.0 of the MCP server: build_context, to ask a whole folder and get only the context that is needed, with measured token counts. You can also try it on the web.
Reads any document in one call
Unlocks it, runs local OCR on scans, identifies the type and extracts the key fields with their page and evidence.
Content is data, not orders
Text a person cannot see (white, tiny, off the page) is flagged before the document. A PDF does not get to instruct your agent.
Only inside the folder you choose
Every path is checked after resolving symlinks. Nothing is overwritten without overwrite: true, and there is a read-only mode.
What you can ask
Almost any file goes in. What your AI already understands comes out.
Any cloud suite can promise to delete your file. We do something different: we never receive it, and you can check.
Open the inspector (F12 → Network) while you process a file. No request carries your document, because there is no server to receive it.
Four ways to verify it0 requests · 0 B transferred
while you process your file
IDs, names, emails and phones become consistent placeholders before they reach the AI. Paste the answer back and the real data returns. Nothing is destroyed.
Patient Juan Pérez, ID 1.020.345.678, tel 310 555 0199
Patient [NOMBRE_1], ID [DOCUMENTO_1], tel [TELEFONO_1]
Restore the real data
Turn off WiFi and keep working. If your file were uploaded anywhere, this would be impossible.
The site ships a CSP header that blocks the code's connections (fetch, XHR, WebSocket, beacons) to any other domain. You can see it in DevTools.
Content-Security-Policy: connect-src 'self'MCP server, CLI, Python and n8n run on your machine. Claude Code or Cursor read your documents without them leaving it.
claude mcp add sirdas -- npx -y @sirdaspdf/mcp@0.4.3For developers No size limits, no watermarks, no account.
Drop the file. It opens in your browser's memory.
Choose what to do. The work runs on your processor.
Download the result. Close the tab and nothing remains.
Each one opens straight away. No queue, no sign-up, no daily limit.
Text, tables and data ready for your model
Smaller files, same quality
Merge, split, reorder, delete and extract pages
Page numbers and watermarks
Add and remove passwords
Images to PDF and back
New niche
Drop a folder of PDFs and get a JSONL ready to fine-tune a model or feed your RAG: redacted, deduplicated and split by document. It is the chain no cloud parser can offer with client documents.
Split by document
Never by chunk. Otherwise the model already saw the template and the metric lies.
MinHash + LSH deduplication
Repeated text gets memorized instead of learned. Threshold 0.8, verified with exact Jaccard.
Personal data out, and reversible
A model repeats what it sees. Redaction is on by default and the mapping stays with you.
Out in the format your trainer expects
Cloud suites promise to delete it within an hour or two. Here there is nothing to delete.
| Tool | Uploads your file? | How long do they keep it? | Works offline? |
|---|---|---|---|
| Cloud suites, web version (iLovePDF, PDF24, Smallpdf) | Yes, to their servers | Between 1 and 2 hours, per each policy | No |
| Sırdaş | No. It never leaves your browser | Nothing to keep | Yes |
Source: each service's public privacy policy, reviewed in September 2026. The table compares their web versions: iLovePDF Desktop, PDF24 Creator and the Smallpdf Windows app process on the device, but have to be installed.
Articles on preparing documents for AI without handing them to anyone. With citations, with what does not work, and with what we still cannot do.
From PDF to JSONL: extract the text, redact, drop duplicates, split by document and export in the format your trainer expects. All on your machine.
9 minDuplicates, contamination, leakage through chunking and personal data: four documented mistakes, how to fix each, and why the third one fools your metric.
11 minWhat the public policies of the most used PDF sites say about how long they keep your file, and how to check yourself whether a tool uploads it.
8 minYour documents tell your life: your health, your contracts, your home.
You shouldn't hand them to an unknown server just to resize them.
That's why nothing is uploaded here. Not today, not when you pay, not ever.
Today, the same tools on your own machines. Being built with pilot companies: team console, shared templates and a data processing agreement.
See the business offerIt takes less time than reading this sentence. And your file stays where it was.
Try it now