MCP · CLI · Python · n8n · TypeScript · Docker
The document layer for your AI agents
PDFs, scans, Word, Excel, PowerPoint, email and photos turned into data an agent can cite: document type, fields with page and position, tables, forms, chunks and datasets. On your machine: no uploads, no credits, no queue.
Using an AI agent?
Copy these instructions into Claude, ChatGPT or Cursor. The agent reads our Markdown docs, installs itself and tells you what it can do.
Are you an agent?
This whole site has a plain-text version:
- /llms.txt — index
- /llms-full.txt — everything in one file
- /skills/…/SKILL.md — Agent Skill
- /docs/agents/*.md — reference
Install it where you already work
One line. Then ask it to “read invoice.pdf”.
claude mcp add sirdas -- npx -y @sirdaspdf/mcp
One call does it all
read_document unlocks, runs OCR when needed, identifies, extracts with citations, anonymizes, chunks and tells you what to offer next. This is what it returns:
{
"status": "ok",
"format": "pdf",
"document_type": "factura",
"steps": ["Extracted text from 2 page(s).", "Identified as factura (match 1).", "Extracted 8 field(s) with page citations.", "Replaced 3 personal data item(s) with placeholders."],
"fields": [
{ "name": "total", "value": "1.428.000", "page": 2,
"evidence": "Total a pagar: $ 1.428.000", "bbox": [0.08, 0.61, 0.43, 0.63] },
{ "name": "cufe", "value": "8f3a2b1c…", "page": 1, "bbox": [0.08, 0.22, 0.71, 0.24] }
],
"chunks": [{ "id": "fnv1a:9c1e…", "pages": [1, 2], "section": ["Factura"], "tokens": 412, "text": "…" }],
"next_steps": [{ "tool": "pdf_tables", "why": "Export the invoice line items to CSV." }],
"bytes_sent": 0
}Almost any file goes in
- PDF and scanned PDF
- PNG and JPG
- Word (.docx), .odt, .rtf
- Excel (.xlsx), .ods, CSV
- PowerPoint (.pptx), .odp
- EPUB and HTML
- .eml emails
- JSON, Markdown, text
Out in the format your stack already uses
- Chat (OpenAI, TRL)
- Alpaca (Axolotl, LLaMA-Factory, Unsloth)
- ShareGPT
- LangChain Documents
- LlamaIndex TextNodes
- Embeddings (id + text + metadata)
- Hugging Face dataset card
- train / validation / test by document
Why local
Provable privacy
Medical records, contracts and payroll never leave the machine. The MCP server opens no network connections.
No credits, no queue
Cloud parsers charge per page. Here 100,000 pages cost the same as one.
Every value is checkable
Each field carries its source line and its box on the page: the agent cites, it doesn't guess.
What it doesn't do yet
Reading handwriting, turning charts into data or detecting a pen-ticked box on a scan needs a vision model, and we don't claim it. Word, Excel and PowerPoint are read for content, not re-rendered with their layout.
Full documentation
- Servidor MCPFor Claude Code, Claude Desktop, Cursor, VS Code and any MCP client: 27 local tools.
- CLITerminal, scripts and CI. --json output on every command.
- Pythonpip install sirdas, with LangChain and LlamaIndex loaders.
- n8nCommunity node for workflows: email → invoice → spreadsheet.
- TypeScriptThe @sirdaspdf/core library, in Node and in the browser.
- Agent SkillTeaches an agent when and how to use Sırdaş.
- Privacy modelWhat guarantees an agent can repeat to its users.