mineru-selfhosted-mcp 0.1.17

MCP bridge for a self-hosted MinerU API

mineru-selfhosted-mcp

mineru-selfhosted-mcp exposes a self-hosted MinerU service as an MCP server.

It is designed for setups where:

• MinerU is deployed on a remote GPU server
• Claude Desktop / Cursor / other MCP clients run elsewhere
• The client should only run a lightweight MCP bridge locally

Environment variables

• MINERU_BASE_URL: Optional. Base URL of your self-hosted MinerU API. Defaults to the public FRP endpoint http://42.51.34.112:8191
• MINERU_API_KEY: Optional token sent to the MinerU API as Authorization: Bearer <token>
• MINERU_TIMEOUT: Optional request timeout in seconds. Default: 1800
• MINERU_TRUST_ENV: Optional. Set to true only if you want the bridge to inherit local proxy variables. Default: disabled.
• MINERU_LOG_DIR: Optional. Directory used by clean_logs. Defaults to ~/.mineru-selfhosted-mcp/logs
• MINERU_DOWNLOAD_PUBLISH_DIR: Optional. Directory used for published download artifacts. Defaults to /data/mineru_output/_published
• MINERU_DOWNLOAD_BASE_URL: Optional. Public base URL for published artifacts. Defaults to <MINERU_BASE_URL>/downloads
• MINERU_ZIP_TTL_SECONDS: Optional. Temporary ZIP lifetime before cleanup. Defaults to 10800 seconds (3 hours).
• MINERU_MCP_TRANSPORT: Optional. stdio for local clients or http for a Linux-hosted remote MCP service. Default: stdio
• MINERU_MCP_HOST: Optional bind host for HTTP MCP mode. Default: 0.0.0.0
• MINERU_MCP_PORT: Optional bind port for HTTP MCP mode. Default: 8001
• MINERU_UPLOAD_ROOT: Optional. Directory used for staged uploaded source files. Defaults to /data/mineru_uploads

If you are running on the Linux host itself, the local proxy listens on http://127.0.0.1:8192. Remote MCP clients should use the public FRP endpoint on 8191.

Connection modes

1. Windows local MCP via stdio

Use this when Claude Code, Cursor, or Claude Desktop runs on Windows and only the MinerU API lives on Linux.

• MCP runs locally on Windows
• MinerU API runs remotely on Linux
• output_dir, artifact_paths, and zip_path are Windows-local
• *_download_url is convenience-only

2. Linux remote MCP via HTTP

Use this when you want one shared MCP service for many clients and durable public artifact URLs.

• MCP runs on Linux
• MinerU API runs on Linux
• /downloads is served from the same Linux filesystem
• Windows clients connect with type: "http"
• zip_download_url and other *_download_url fields become first-class outputs

Exposed tools

• mineru_health: check the remote MinerU API
• parse_document: parse a single local file through the remote MinerU API
• parse_documents: parse one or more local files through the remote MinerU API
• parse_directory: parse all matching PDFs in a directory through the remote MinerU API
• create_upload_session: create a staged upload session for a client-side file
• append_upload_chunk: append a base64 chunk into an upload session
• finalize_upload: validate and finalize an uploaded staged file
• parse_staged_file: parse a finalized staged file
• parse_inline_document: upload a base64-encoded local file and parse it in one call
• parse_url_document: fetch a remote URL, stage it on Linux, and parse it in one call
• cleanup_staged_file: remove a staged upload session
• get_ocr_languages: list common OCR language codes supported by MinerU
• clean_logs: remove local MCP log files older than a chosen number of days

Parsing result metadata

Parsing tools also return:

• download_url
• zip_path
• zip_download_url
• elapsed_seconds
• file_count
• completed_count
• failed_count
• progress

progress is a stable completion summary for the current call rather than a live streaming progress feed.

parse_document, parse_documents, parse_directory, and submit_parse_task require local absolute file paths plus an explicit output_dir. They default to return_md=true, return_images=true, formula_enable=true, table_enable=true, parse_method="auto", inline_content=false, and response_format_zip=true.

ZIP results are cleaned before publication: only Markdown files and images/ contents are retained. middle_json, model_output, .cache/, raw model traces, and other intermediate files are removed from the downloadable ZIP. Temporary ZIPs are scheduled for deletion after 3 hours by default.

For batch calls with more than 3 files, inline_content is forced to false to protect MCP context size.

Notes

• To get only middle_json, set middle_json_only=true in parsing tools.
• middle_json_only=true automatically disables markdown in the MCP request wrapper.
• For Windows -> Linux file transfer through MCP, use create_upload_session, append_upload_chunk, finalize_upload, then parse_staged_file.
• For simpler one-shot flows, use parse_inline_document for local bytes or parse_url_document for remote links.

Example MCP config

{
    "mcpServers": {
      "mineru-selfhosted": {
        "command": "uvx",
        "args": ["-y", "mineru-selfhosted-mcp"],
        "env": {
          "MINERU_API_KEY": "your_token",
          "MINERU_BASE_URL": "http://42.51.34.112:8191",
          "MINERU_DOWNLOAD_BASE_URL": "http://42.51.34.112:8191/downloads"
        }
      }
  }
}

Example remote HTTP MCP server

Run this on the Linux host:

MINERU_API_KEY=<TOKEN> \
MINERU_BASE_URL=http://127.0.0.1:8192 \
MINERU_MCP_TRANSPORT=http \
MINERU_MCP_HOST=0.0.0.0 \
MINERU_MCP_PORT=8001 \
MINERU_DOWNLOAD_PUBLISH_DIR=/data/mineru_output/_published \
MINERU_DOWNLOAD_BASE_URL=http://42.51.34.112:8191/downloads \
uvx -y mineru-selfhosted-mcp

Then expose the MCP endpoint through nginx or another reverse proxy, for example:

• MCP endpoint: http://42.51.34.112:8191/mcp/
• downloads base: http://42.51.34.112:8191/downloads/

Example Windows Claude Code config for remote HTTP MCP

{
  "mcpServers": {
    "mineru-selfhosted": {
      "type": "http",
      "url": "http://42.51.34.112:8191/mcp/"
    }
  }
}