DeepSeek Harness × Baidu Netdisk: Read-Only Directory Audit

An unofficial, community-maintained minimal project: let DeepSeek Harness read your own netdisk through Baidu Netdisk's official MCP, recursively organize large tutorial directories, while blocking download and write operations by default.

This repository does not contain anyone's Access Token, real netdisk directories, file listings, or audit reports, nor does it install, sync, or upload netdisk files.

Why the bridge is needed

Baidu Netdisk officially provides the remote SSE address https://mcp-pan.baidu.com/sse. DeepSeek Harness's generic MCP client supports stdio and Streamable HTTP; this project provides a small SSE → stdio adapter and uses an explicit read-only allowlist in the adapter.

DeepSeek Harness
      │ stdio
      ▼
本地只读桥接器
      │ 官方 SSE + 用户自己的 Access Token
      ▼
百度网盘官方 MCP

Only these tools are registered in the Harness:

• file_list

• file_doc_list

• file_image_list

• file_video_list

• file_meta

• file_keyword_search

• file_semantics_search

• user_info

• get_quota

Tools for upload, download, sharing, deletion, moving, copying, renaming, and creating directories are not exposed to the model.

Related MCP server: CodeAudit MCP

Prerequisites

• Windows 10/11 (the current script is verified on Windows; the bridge itself is cross-platform Node.js)

• Node.js 20 or higher

• An officially launched DeepSeek Harness

• An Access Token obtained through Baidu Netdisk's official process

Official references:

• Baidu Netdisk MCP

• DeepSeek Harness MCP client

Installation

git clone https://github.com/MarenV-narrative/deepseek-harness-baidu-netdisk-audit.git
Set-Location '.\deepseek-harness-baidu-netdisk-audit'
npm install

npm install only installs the publicly declared MCP SDK dependencies; it does not install or overwrite DeepSeek Harness.

1. Securely save the Access Token

Complete the authorization following the "Getting Started" section of Baidu's official README. Do not write the token into source code, Git, command-line arguments, or chat content.

On Windows, use the hidden-input script:

& '.\scripts\set-token-user.ps1'

It saves the token as the BAIDU_NETDISK_ACCESS_TOKEN environment variable for the current Windows user. After saving, close and reopen the Harness.

2. Generate the Harness configuration snippet

npm run generate-config

The script generates config/harness-cordis-entry.generated.yml, which contains only the absolute paths of the local Node and the project, with no token. Append the entire patch operation to:

%USERPROFILE%\.dsh\profiles\web\cordis.patch.yml

Back up the original file first; do not overwrite an existing patch. After restarting the Harness, the mcp__baidu-netdisk__... tools should appear on the model side.

You can run a connection test that does not read private filenames:

npm run smoke

It only fetches the bridge's public tool list and confirms that no download or write tools are present.

3. Specify your own netdisk directory

Edit config/audit.json:

{
  "target_dir": "/网络安全教程",
  "dir_encoding": "urlencoded",
  "max_pages_per_directory": 10000,
  "document_content_preview_chars": 2000
}

You can also override the target for this test only:

& '.\scripts\run-readonly-test.ps1' --target '/网络安全教程'

4. Minimal read-only verification

& '.\scripts\run-readonly-test.ps1'

This test will:

1. Use file_list to paginate the target directory;

2. Recursively obtain the subdirectory tree;

3. Select one document and use file_meta to check its metadata, content, and abstract;

4. Write the private results to the local inventory/ and reports/.

These two output directories are ignored by Git by default. The test script makes no download, upload, or write calls to the netdisk.

5. Organize tutorials in the Harness

Copy the prompt from prompts/course-audit.zh.md and replace the target directory with your own path.

If content or abstract is empty, it usually means Baidu has not yet generated the parsing result, not that the file is empty. First infer from the path, filename, extension, size, and adjacent courses; only build an on-demand check queue for low-confidence files that affect learning order or installation safety.

This project deliberately does not implement a download tool. If on-demand downloads are added later, they should be placed in a separate tool that requires manual confirmation and limits single-file size and local disk output.

Privacy and security

• The remote address is fixed to Baidu's official domain, avoiding environment variables forwarding the token to arbitrary servers.

• The token is passed to the bridge subprocess only through environment variables.

• Error messages attempt to mask the token.

• Read-only permissions are enforced by the code allowlist, not just by prompts.

• inventory/ and reports/ may contain private filenames, paths, summaries, and content previews; do not commit them.

• Run npm run privacy-check to check for common token shapes, Windows user paths, and known private paths.

Compatibility notes

This project was initially verified with: DeepSeek Harness 0.1.0-rc.6, MCP SDK 1.30.0, Windows 11. If upstream changes the transport method, tool names, or configuration format, refer first to the official documentation of Baidu Netdisk and DeepSeek Harness.

Disclaimer

This project is not an official Baidu or DeepSeek project and has no affiliation, endorsement, or partnership with them. Users should comply with Baidu Netdisk's terms of service and access only data they own or are authorized to access.