dsh-vidfetch

Give your DeepSeek Harness agent the ability to download a video straight from a URL.

What it does

Registers a single tool, fetch_video, that the model can call to download the main video from any URL.

• Downloads the best available audio+video stream into the directory the agent specifies.
• Returns ok plus the list of files it wrote. Or a short error cause if the URL is unsupported.
• No configuration file or extra setup on the agent side: just point dir somewhere writable.

Installation

dsh plugin --profile web add dsh-vidfetch

From this repository:

git clone https://github.com/anaksunamu/dsh-vidfetch.git 
npm install --prefix ./dsh-vidfetch
npm run build --prefix ./dsh-vidfetc
dsh plugin --profile web add ./dsh-vidfetch

Usage

The model calls fetch_video with two required parameters:

Example call the agent makes:

{ "url": "https://example.com/watch/abc123", "dir": "~/downloads" }

Output:

{ "message": "ok", "files": ["downloads/video.mp4"] }

On failure, message contains a short error cause and files is empty:

{ "message": "Unsupported URL", "files": [] }

Requirements

• DSH: tested against the 0.1.1-rc.x line.
• Node: ^22.19.0 || >=24.0.0.
• yt-dlp: provided by dependency.

Permissions & network

Clear the risk explicitly before trusting an agent tool:

• Outbound network: each call fetches whatever URL the model passes. The plugin sends no telemetry and contacts no external service of its own.
• Filesystem write: files land in the dir the model supplies (default: current working directory if empty). The model decides the path — point dir at a sandboxed or scratch folder if you don't want arbitrary writes.
• Subprocess: runs a yt-dlp process per call; it inherits the host's environment.
• Untrusted content warning: videos fetched from random URLs are untrusted files. Don't pipe them straight into execution pipelines.
• No API key required by this plugin.

License & listing

• License: MIT.
• Not affiliated with DeepSeek AI.
• Compatibility notes above are for DSH preview 0.1.1-rc.x.