Mineru Extract

MinerU Extract (official API)

Use MinerU as an upstream “content normalizer”: submit a URL to MinerU, poll for completion, download the result zip, and extract the main Markdown.

Quick start (MCP-aligned)

We align to the MinerU MCP mental model, but we do not run an MCP server.

• Primary script (MCP-style): scripts/mineru_parse_documents.py
• Input: --file-sources (comma/newline-separated)
• Output: JSON contract on stdout: { ok, items, errors }
• Low-level script (single URL): scripts/mineru_extract.py

Auth:

• Set MINERU_TOKEN (Bearer token from mineru.net)

Default model heuristic:

• URLs ending with .pdf/.doc/.ppt/.png/.jpg → pipeline
• Otherwise → MinerU-HTML (best for HTML pages like WeChat articles)

1) Configure token (skill-local)

Put secrets in skill root .env (do not paste into chat outputs):

# /home/node/.openclaw/workspace/skills/mineru-extract/.env
MINERU_TOKEN=***
MINERU_API_BASE=https://mineru.net

2) Parse URL(s) → Markdown (recommended)

MCP-style wrapper (returns JSON, optionally includes markdown text):

python3 skills/mineru-extract/scripts/mineru_parse_documents.py \
  --file-sources "\n" \
  --language ch \
  --enable-ocr \
  --model-version MinerU-HTML

If you want the markdown content inline in the JSON (can be large):

python3 skills/mineru-extract/scripts/mineru_parse_documents.py \
  --file-sources "" \
  --model-version MinerU-HTML \
  --emit-markdown --max-chars 20000

Low-level (single URL, print markdown to stdout):

python3 skills/mineru-extract/scripts/mineru_extract.py "" --model MinerU-HTML --print > /tmp/out.md

Output

The script always downloads + extracts the MinerU result zip to:

/home/node/.openclaw/workspace/mineru//

It writes:

• result.zip
• extracted files (Markdown + JSON + assets)

It prints a JSON summary to stderr with paths:

• task_id, full_zip_url, out_dir, markdown_path

Parameters (common)

• --model: pipeline | vlm | MinerU-HTML (HTML requires MinerU-HTML)
• --ocr/--no-ocr: enable OCR (effective for pipeline/vlm)
• --table/--no-table: table recognition
• --formula/--no-formula: formula recognition
• --language ch|en|...
• --page-ranges "2,4-6" (non-HTML)
• --timeout 600 / --poll-interval 2

Failure modes & fallbacks

• MinerU may fail to fetch some URLs (anti-bot / geo / login).
• Fallback: provide an HTML file or a PDF/long screenshot; then implement “upload + parse” flow with MinerU batch upload endpoints.
• Always report the failing URL + MinerU err_msg and keep an original-source link in outputs.

References

• MinerU API docs: https://mineru.net/apiManage/docs
• MinerU output files: https://opendatalab.github.io/MinerU/reference/output_files/

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: blessonism
• Source: blessonism/openclaw-skills
• License: MIT
• Homepage: https://github.com/openclaw/openclaw

Install and usage instructions live in the source repository linked above.