okra-public-docs（秋葵公共文档）

OkraPDF Public Documents

Pre-extracted public document corpora queryable via MCP. No upload, no waiting — documents are already parsed and indexed. Just pass an ID and start asking questions.

Available Channels

Setup

Arxiv papers (authenticated MCP)

Add to ~/.claude/mcp.json (Claude Code) or .cursor/mcp.json (Cursor):

{
  "mcpServers": {
    "okra-pdf": {
      "type": "url",
      "url": "https://api.okrapdf.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_KEY" }
    }
  }
}

Get a free API key at okrapdf.com (Settings > API Keys).

SEC filings (zero-auth MCP)

{
  "mcpServers": {
    "okra-sec": {
      "type": "url",
      "url": "https://mcp.okrapdf.com/mcp"
    }
  }
}

No API key, no signup. Restart your agent after adding.

Arxiv Papers

400+ recent AI research papers parsed with Docling OCR on GPU — tables, equations, figures, and full text preserved as structured markdown.

Read a paper

read_document(document_id: "arxiv:2603.26653")
read_document(document_id: "arxiv:2603.26653", pages: "1-5")
read_document(document_id: "https://arxiv.org/pdf/2603.26653")

No upload needed — papers are pre-indexed as public sources. Just pass the arxiv ID.

Ask questions

ask_document(document_id: "arxiv:2603.26653", question: "What is the main contribution?")
ask_document(document_id: "arxiv:2603.26653", question: "What were the benchmark results on MMLU?")
ask_document(document_id: "arxiv:2603.18272", question: "How is retrieval-augmented experience used?")

Returns answer with page citations.

Extract structured data

extract_data(
  document_id: "arxiv:2603.26653",
  prompt: "Extract all benchmark results with model names, dataset names, and scores",
  json_schema: {
    "type": "object",
    "properties": {
      "benchmarks": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "model": {"type": "string"},
            "dataset": {"type": "string"},
            "metric": {"type": "string"},
            "score": {"type": "number"}
          }
        }
      }
    }
  }
)

Literature survey workflow

# Read abstracts from several papers
read_document(document_id: "arxiv:2603.26499", pages: "1")
read_document(document_id: "arxiv:2603.26266", pages: "1")

# Ask targeted questions
ask_document(document_id: "arxiv:2603.26499", question: "What bottlenecks in AI research does this address?")

# Same question across papers for comparison
ask_document(document_id: "arxiv:2603.18272", question: "How does this handle multi-agent coordination?")
ask_document(document_id: "arxiv:2603.07379", question: "How does this handle multi-agent coordination?")

Discover papers

Semantic Scholar (free, no key needed for basic use):

curl -s "https://api.semanticscholar.org/graph/v1/paper/search?query=agentic+RAG&year=2026&fields=externalIds,title,citationCount&limit=10" \
  | jq '.data[] | {arxiv: .externalIds.ArXiv, title, citations: .citationCount}'

Arxiv RSS feeds (same feeds used to build the collection):

https://rss.arxiv.org/rss/cs.AI    # Artificial Intelligence
https://rss.arxiv.org/rss/cs.CL    # Computation and Language (NLP)
https://rss.arxiv.org/rss/cs.LG    # Machine Learning

Papers With Code:

curl -s "https://paperswithcode.com/api/v1/papers/?q=agentic+RAG&items_per_page=5" | jq '.results[] | {title, arxiv_id}'

Current snapshot

411 papers from cs.AI (~200), cs.CL (~100), cs.LG (~200). Full manifest in papers.json.

If a paper isn't found, upload it yourself with upload_document.

Tips

• Use arxiv:XXXX.XXXXX format (not full URL) for cleaner queries
• pages: "1" reads just the abstract/intro quickly
• For survey papers (50+ pages), use ask_document instead of reading everything
• extract_data with JSON schemas is ideal for pulling benchmark tables

SEC Filings

Pre-extracted SEC 10-K and 10-Q filings. No API key, no signup, completely free.

Available tools

Browse filings

read_filing_index()
read_filing_index(ticker: "NVDA")
read_filing_index(ticker: "AAPL", filing_type: "10-K")

Always start here to see what's available.

Ask questions (single company)

ask_question(question: "What was NVIDIA's data center revenue?", tickers: ["NVDA"])
ask_question(question: "List all risk factors related to AI regulation", tickers: ["MSFT"])
ask_question(question: "What are the outstanding debt obligations?", tickers: ["TSLA"], filing: "10-k-2024")

Cross-company comparison (up to 10 tickers)

ask_question(
  question: "Compare R&D spending as a percentage of revenue",
  tickers: ["AAPL", "MSFT", "GOOGL", "NVDA", "META", "AMZN", "TSLA"]
)

ask_question(
  question: "Which company has the highest gross margin?",
  tickers: ["AAPL", "MSFT", "GOOGL"]
)

ask_question(
  question: "Summarize each company's AI strategy",
  tickers: ["NVDA", "AMD", "INTC"]
)

Fans out to each company's filing in parallel, then synthesizes a cross-company answer.

Read full filing text

read_filing_contents(ticker: "TSLA", filing: "10-k-2024")

Filing slug formats (all equivalent): 10-k-2024, 10-K/2024, 2024-10K.

Extraction quality audit

get_verification_summary(document_id: "doc-xxx")
get_verification_summary(document_id: "doc-xxx", status: "needs_review")
verify_pages(document_id: "doc-xxx", action: "approve", confidence_above: 0.9)
verify_pages(document_id: "doc-xxx", action: "flag", pages: [67], reason: "Table has merged cells")

Available companies

Mag7: AAPL, MSFT, GOOGL, AMZN, NVDA, META, TSLA

FinanceBench: ~80 companies including major banks (JPM, BAC, GS), pharma (PFE, JNJ, ABBV), industrials (GE, MMM, CAT), and more.

Use read_filing_index() to browse the full catalog. New filings added as published.

Tips

• Start with read_filing_index before querying
• ask_question with multiple tickers is the fastest way to compare — no need to read each filing
• Cross-company queries work best with clear, quantitative questions
• Verification tools require document_id (not ticker) — get it from other tool responses