stanford-edgar-parser 0.1.2

Layout-faithful SEC EDGAR filing parser from the Stanford EDGAR Filings Dataset.

Stanford EDGAR Filings Dataset Parser

The Stanford EDGAR Filings Dataset (SEFD) is a 550B-token reconstruction of 18.5M SEC filings into layout-faithful MultiMarkdown for LLM pretraining, financial reasoning, document understanding, and evaluation.

SEFD reverse-engineers EDGAR's heterogeneous disclosure formats into a token-efficient representation that preserves financial tables, indentation, merged headers, numeric signs, currency and percent symbols, document hierarchy, and other layout cues that carry financial meaning. Internal validation shows our rule-based reconstruction methodology achieves greater than 99% structural and semantic accuracy on sampled outputs.

The software routes and parses the full EDGAR source-format surface, including legacy fixed-width text, tag-soup HTML, SGML wrappers, XML submissions, and PDF attachments, with specialized reconstruction for more than 30 SEC XML schemas, including Forms 3, 4, 5, D, 13D/G, N-PX, N-PORT, N-CEN, 13F, 144, ATS-N, 1-A/K/Z, C, MA, TA, X-17A-5, 24F-2NT, ABS-EE, and related amendments, withdrawals, and corrections. PDF attachments are parsed with Mistral OCR 3.

Our first public release, SEFD-v1, is a 152B-token dataset covering filings from January 2022 through June 2025.

• Paper: The Stanford EDGAR Filings Dataset
• Dataset: SEFD-v1 on Hugging Face

Install

From PyPI:

pip install stanford-edgar-parser

Agent-friendly install alias:

pip install "stanford-edgar-parser[ai]"

Or directly from GitHub:

pip install "stanford-edgar-parser[ai] @ git+https://github.com/Stanford-Advanced-FinTech-Lab-SAFTL/stanford-edgar-filings-dataset.git"

Usage

Parse a local filing and convert tables to MultiMarkdown:

stanford-edgar-parser path/to/filing.txt --to_mmd

or:

python -m stanford_edgar_parser path/to/filing.txt --to_mmd

Optional rendering helpers:

node multimarkdown.js path/to/file.md > file.html
node html-to-pdf.mjs file.html file.pdf

Agent Setup

The [ai] extra is an agent-facing install alias. It currently uses the same runtime dependencies as the base package, but gives agent clients a stable target compatible with package patterns such as edgartools[ai].

Codex

Install the Codex skill:

uvx --from "stanford-edgar-parser[ai]" stanford-edgar-install-skill --target codex --overwrite

Add the MCP server to ~/.codex/config.toml:

[mcp_servers.stanford_edgar_parser]
command = "uvx"
args = ["--from", "stanford-edgar-parser[ai]", "stanford-edgar-mcp"]
startup_timeout_sec = 120

For full repo-local render/review tools, point Codex at a clone instead:

[mcp_servers.stanford_edgar_parser]
command = "python"
args = ["-m", "stanford_edgar_parser.mcp_server"]
cwd = "/path/to/stanford-edgar-filings-dataset"
startup_timeout_sec = 120

Claude Code

Install the Claude skill:

uvx --from "stanford-edgar-parser[ai]" stanford-edgar-install-skill --target claude --overwrite

For project-local MCP, add this .mcp.json to your project or use the one included in this repository:

{
  "mcpServers": {
    "stanford-edgar-parser": {
      "command": "uvx",
      "args": ["--from", "stanford-edgar-parser[ai]", "stanford-edgar-mcp"]
    }
  }
}

For repo-local render/review tools, use a clone-backed .mcp.json:

{
  "mcpServers": {
    "stanford-edgar-parser": {
      "command": "python",
      "args": ["-m", "stanford_edgar_parser.mcp_server"],
      "cwd": "/path/to/stanford-edgar-filings-dataset"
    }
  }
}

Claude Desktop

Add this to your Claude Desktop MCP config:

{
  "mcpServers": {
    "stanford-edgar-parser": {
      "command": "uvx",
      "args": ["--from", "stanford-edgar-parser[ai]", "stanford-edgar-mcp"]
    }
  }
}

Python Skill Installer

The same bundled skills can also be installed from Python:

from stanford_edgar_parser.ai import install_skill

install_skill(targets=("codex", "claude"), overwrite=True)

GitHub MCP Install

Use the GitHub source directly before a PyPI release:

[mcp_servers.stanford_edgar_parser]
command = "uvx"
args = [
  "--from",
  "stanford-edgar-parser[ai] @ git+https://github.com/Stanford-Advanced-FinTech-Lab-SAFTL/stanford-edgar-filings-dataset.git",
  "stanford-edgar-mcp"
]
startup_timeout_sec = 120

Package-installed MCP always exposes parse_filing. Repo-local rendering and review tools are exposed when the server runs from a full clone containing multimarkdown.js, html-to-pdf.mjs, and tools/.

MCP Tools

The MCP server exposes:

• parse_filing: parse local SEC submissions to Markdown or MultiMarkdown
• render_markdown: render parsed Markdown to PDF when running from a full repo clone
• check_showcase_tables: run static table-integrity checks when running from a full repo clone
• review_snippet: build raw-browser-vs-parsed review snippets when running from a full repo clone

Citation

@article{bettencourt2026stanfordedgar,
  title={The Stanford EDGAR Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data},
  author={Bettencourt, Nick and Ding, Xiaowei and Giesecke, Kay},
  journal={arXiv preprint arXiv:2606.18192},
  year={2026},
  url={https://arxiv.org/abs/2606.18192}
}