DeepSeek Visionary — Claude Code MCP server: web vision for agents
Lets coding agents look at images through DeepSeek's web vision model, with a browser login instead of an API key.
Open on GitHubWhat DeepSeek Visionary does
This MCP server exposes DeepSeek's web vision model as a tool that agents call over stdio. It works with any host that speaks the protocol: Claude, Codex, Cursor, Zed, GitHub Copilot, Cline, Windsurf.
The core is a single Rust binary called visionary-server, a rewrite of an earlier Python project. Run it in CLI mode for vision, ocr, status, login, logout, skill, init or doctor; the mcp-stdio subcommand starts the MCP service. On first login a browser window opens at chat.deepseek.com, and the resulting token lands in ~/.deepseek-visionary/config.json, shared by the CLI, the MCP server and the DeepSeek Harness plugin. A manual path exists too: copy the token from Local Storage in DevTools into the user_token field of that file.
Install the binary before wiring anything up, or the host has nothing to spawn. Hooking an agent in takes one line, for example visionary-server init claude, and --dry-run shows the config before it is written. Zed users get a separate extension shell that fetches and caches the binary.
What it helps with
- read text and layout from a screenshot without an API key
- feed images to an agent that handles text only
- attach vision to Cursor or Codex with a single
initcall - keep the DeepSeek session across CLI and MCP runs
- install the bundled calling skill into an agent's skills folder
Example prompts for the agent
- Describe what is in
screenshot.pngusingdeepseek_vision - Run OCR on
img.pngand return the table text - Check the DeepSeek Visionary login status
How to install DeepSeek Visionary
Install
The MCP registry has no launch command for it: see the repository README
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about DeepSeek Visionary
How do I install DeepSeek Visionary in Claude Code?
Get the binary first, for example with npm install -g @xlight-oss/visionary-server. Then run visionary-server login and add the server with visionary-server init claude.
What does DeepSeek Visionary do?
It hands the agent an image-recognition tool backed by DeepSeek's web model. Authorization happens in a browser window, so no API key is involved.
Do I need a DeepSeek API key?
No. A browser opens, you sign in at chat.deepseek.com, and the server captures the token and session data itself, storing them in ~/.deepseek-visionary/config.json.
