deepseek-eyes MCP
An MCP server that gives text models sight: it takes a clipboard image, sends it to Qwen-VL and returns a text description.
- 63 ★ on the repository
- Last commit
Extended access
Why this level
- Clipboard images are sent to the external ModelScope service together with an API key.
Install
- Works in
Manual install
git clone https://github.com/Shaohan-He/deepseek-eyes.gitThen a virtual environment, `pip install -e .` and MCP setup with a ModelScope key.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Clipboard images are sent to the external ModelScope service together with an API key.
This is third-party code. Review the repository files before installing.
What it does
Models such as DeepSeek and GLM cannot look at images through the API: a pasted screenshot turns into a file path. `deepseek-eyes` fills the gap. The user copies an image, the agent calls an MCP tool and the server reads the image from the clipboard, sends it to the Qwen-VL model on ModelScope and returns a text description. The text model then works from that description, for example retelling a login screen and listing the fields. The server runs over stdio and is configured in Claude Code and OpenCode with an absolute Python path from the virtual environment and a `MODELSCOPE_API_KEY`. The author cites the free ModelScope quota: up to 2000 calls a day and 500 per model.
Who it is for. Claude Code and OpenCode users with text-only DeepSeek or GLM models who need to show the agent screenshots and mockups.
Good fit when
- An agent on DeepSeek must analyze an error screenshot from the clipboard.
- An interface mockup must be described in words before layout work.
- The free ModelScope quota is preferred to a paid vision model.
- Screenshots are needed without access to foreign vision services.
Not a fit when
- The main model already handles images and a go-between is unnecessary.
- Images must not be sent to the external ModelScope service.
- Exact reading of small text is needed: the model description may miss it.
Example request
I will copy a screenshot to the clipboard. Describe what is on it and list the form fields.MCP
- Transport
- stdio
- Authentication
- API key
Environment variables
MODELSCOPE_API_KEYrequired, secret- ModelScope access token used to call the Qwen-VL model
Limitations
The image goes to the ModelScope server, so confidential screenshots should not pass through it. A key needs a ModelScope account linked to an Alibaba Cloud account, and quota terms are set by the platform and may change. The text model sees only the description, not the picture, so details are lost. The configuration needs an absolute Python path from the environment. Step-by-step installation and the clipboard test are written mainly for Windows.
How to disable. Remove the `deepseek-eyes` entry from the client MCP configuration and the repository folder. Revoke the `MODELSCOPE_API_KEY` in the ModelScope console.
FAQ
How do I install it?
Clone the repository, create an environment and run `pip install -e .`, then add the server to the client MCP configuration.
Where do I get a key?
On ModelScope: the console, access tokens section. The first sign-in asks you to link an Alibaba Cloud account.
How do I test the clipboard?
Copy an image and run `python examples/smoke_test.py`. On success the script reports that the image was saved.
Which clients does it work in?
The README gives settings for Claude Code and OpenCode.
What language is the documentation in?
Chinese with an English version in `README_EN.md`.
Discussion
No questions or reviews yet.
Related
MCP serversMCP server that connects DeepSeek V4.1 Flash and V4 Pro to Claude Code and Codex with images, FIM and the Responses API
An MCP server that gives Claude Code and Codex a cheap DeepSeek helper for classification, summaries and text cleanup.
An MCP server that runs DeepSeek or your own local model as a sub-agent in Claude Code and Codex with its own tool loop.
A skill that gives text-only models sight: an image by path or from a paste becomes structured JSON with transcribed text, layout and semantics
A skill that gives image recognition to models without native vision: it sends the image to a vision model and returns a text description
Local memory for Claude Code, Codex and Cursor Agent: learns from your activity on a Mac and gives agents context with source receipts over MCP
README in short
The README covers one-prompt and manual installation, getting a ModelScope key, the flow diagram, a comparison with similar tools and client configurations.