Text Llm Vision
shaoqiuyuavailable/text-llm-vision · model-adapters
Scene-aware vision routing layer for DeepSeek Harness (dsh): decides which engine/backend an image should go to (chat/UI/table/code) before other vision plugins route it.
★ 2GitHub stars
0Forks
2026-08-17Last updated
PythonLanguage
NOASSERTIONLicense
Key features
- English / KeywordsMCP vision tools (primary path) + a paste-image reverse-proxy fallback for text-only LLMs (DeepSeek, Qwen, Kimi, GLM) used by any AI coding agent (Claude Code / Cline / OpenCode / Codex / Continue / Copilot / Cursor).
- Three-layer designMCP server (describe_image), reverse proxy (paste-image fallback), CLAUDE.md soft rules.
- Triple protocolAnthropic Messages, OpenAI Chat Completions, OpenAI Responses — decoupled from any vendor.
- Vision backendlocal vision model (default Qwen2.5-VL, swappable via config) with Scan/Zoom/Guess three-pass pipeline.
- Zero API cost, fully local/offline, Docker deployable, VS Code control panel.
Install command
dsh plugin --profile web add github:shaoqiuyuavailable/text-llm-vision