MIT 授權的視覺工具箱，讓純文字 agent 具備看圖能力；透過 dsh-vision-toolkit Profile Bundle 原生接入 DeepSeek Harness，CLI 與技能也相容任何可呼叫 shell 的 agent。
DSH 整合
原生執行環境
作者聲明
安全稽核
未稽核
最後核實
2026-08-21
授權條款
MIT
01它能幫你完成什麼？
Give a text-only coding agent (e.g. DeepSeek) working vision: image Q&A, OCR of long screenshots and chat histories
Task-aware image answers, transcribed text, and merged OCR of scrolling pages with speaker/timestamp preservation
Developers running text-only models such as DeepSeek who need image understanding without switching to a multimodal model
Rebuild UI pages, infographics, icons and diagrams from screenshots or sketches into editable code
HTML/CSS reconstructions, transparent PNG extractions, editable SVG, and structured Mermaid/Graphviz representations
Frontend developers and designers who want agent-driven screenshot-to-code / sketch-to-interface restoration
Operate a GUI from screenshots: locate controls, act, and verify the resulting state
Pixel-coordinate grounding of controls and a verify-after-each-action automation loop
Automation engineers who need screen-element grounding (e.g. `ground`/`detect`) for GUI control with a text-only agent
02如何將外掛接入 DeepSeek Harness？
先決條件
- A coding agent already working with a model, including a text-only model such as DeepSeek V4
- A multimodal vision API supporting OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages, plus its base URL, API key, and model name
- Python 3.11+ for `glance`; optional `pillow` for `ground`/`detect`/`crop` and the long-screenshot OCR playbook, `pillow` + `numpy` for `trace`
安裝步驟
- 01
Point it at a vision API — write three env vars (`VISION_API_KEY`, `VISION_BASE_URL`, `VISION_MODEL`) to `~/.config/agent-vision-toolkit/env` (`chmod 600`)
- 02
Put the CLIs on your PATH: `git clone https://github.com/Anionex/agent-vision-toolkit.git` then `export PATH="$PWD/agent-vision-toolkit/bin:$PATH"` (add to your shell profile to persist)
$ git clone https://github.com/Anionex/agent-vision-toolkit.git
- 03
Install the skill so your agent knows the tools exist: `npx skills add Anionex/agent-vision-toolkit --skill vision-skills -a codex -g --copy -y`, or copy `skills/vision-skills/` into your agent's skills directory (e.g. `~/.codex/skills/`) and restart the agent
$ npx skills add Anionex/agent-vision-toolkit --skill vision-skills -a codex -g --copy -y
- 04
Optional seamless integration: follow AGENT_INSTALL.md to install the transparent local proxy (Codex / Claude Code) or single-file native extension/plugin (Pi, Oh My Pi, OpenCode); for native DSH support use the linked `dsh-vision-toolkit` package (clone with `--recurse-submodules` or run `git submodule update --init --recursive`)
$ git submodule update --init --recursive
驗證整合成功
- After installing the optional integration and restarting the agent, paste an image directly or let the model call its built-in image tool
03DSH 整合程度與能力邊界
Native DSH Profile Bundle via the linked `dsh-vision-toolkit` package (tracked as a Git submodule); the standalone CLIs and `vision-skills` skill also run in any shell-capable agent
glance — intent-aware image Q&A and OCR
an image file plus a question (or `--ocr`)→task-aware answers or transcribed text, focused on the current step's intent instead of a generic description
sends the image and focus hint to the configured external multimodal vision APIground / detect — pixel-coordinate localization
an image plus a target description (optionally `--region`)→bounding boxes in original pixel coordinates; numbered element inventories with exact visible text
sends the image to the configured external multimodal vision APItrace / crop — deterministic SVG fitting and region cropping
a flat, high-contrast graphic (or a pixel box for `crop`)→editable SVG primitives (`<circle>`, `<line>`, `<polyline>`, `<polygon>`) fitted locally; cropped region files for reuse
writes output files (SVG / cropped images) to diskSeamless integration — transparent local proxy / native extensions
agent requests containing pasted images or built-in image tool calls→images replaced with task-aware text descriptions before reaching the text-only upstream, with no extra prompting
runs a local proxy at 127.0.0.1:19100 and forwards requests (including the original Authorization header) to your existing upstream
04適合誰？何時不該用？
適合
- Developers running text-only models such as DeepSeek who need image understanding without switching to a multimodal model
- Frontend developers and designers who want agent-driven screenshot-to-code / sketch-to-interface restoration
- Automation engineers who need screen-element grounding (e.g. `ground`/`detect`) for GUI control with a text-only agent
不適合
- This is an image-to-text layer — it doesn't hand vision tokens directly to the text model, and overall visual-task quality is determined jointly by the primary LLM and the multimodal LLM.
05相容性、維護與安全提醒
- This is an image-to-text layer — it doesn't hand vision tokens directly to the text model, and overall visual-task quality is determined jointly by the primary LLM and the multimodal LLM.
- Requires an external multimodal API (OpenAI Chat Completions / Responses / Anthropic Messages) with its own base URL, API key, and model name; images leave your machine to that endpoint.
- The proxy's cache lives only inside its process and is cleared on restart.
MIT · actively maintained (latest release v0.2.0, 2026-08-14)
06常見問題
如何把 agent-vision-toolkit 接入 DeepSeek Harness？
透過配套的 `dsh-vision-toolkit` 套件——它是 DSH Web 與 Headless 設定的原生 Profile Bundle，提供 10 個結構化視覺工具，並附 DSH Credentials、託管隔離執行環境與 Agent 級漸進式工具揭露。它以 Git submodule 形式追蹤：複製時加 `--recurse-submodules`，或在既有儲存庫執行 `git submodule update --init --recursive`。
DSH 整合是原生還是走 MCP？
原生——`dsh-vision-toolkit` 以 Profile Bundle 形式在 DSH 內原生執行，不透過 MCP。另外，基礎工具箱的 CLI 可被任何能呼叫 shell 的 agent 直接使用，不需要任何整合。
需要哪些前置條件？
一個已能搭配模型運作的編程 agent（DeepSeek V4 這類純文字模型即可），以及一個支援 OpenAI Chat Completions / Responses / Anthropic Messages 的多模態視覺 API（含 base URL、API key 與模型名稱）。`glance` 需要 Python 3.11+；`ground`/`detect`/`crop` 需要 `pillow`；`trace` 需要 `pillow` + `numpy`。
我的圖片和資料會傳到哪裡？
圖片內容與任務聚焦提示會傳送到你在 `VISION_BASE_URL` 設定的多模態視覺 API。選用的本機代理執行於 127.0.0.1:19100，僅原封不動轉發你原有的上游 Authorization 標頭，不會另外儲存上游 API key。
有哪些限制要注意？
它是圖像轉文字層——視覺 token 不會直接進入文字模型；視覺任務品質由主模型與多模態模型共同決定；代理快取只存在於處理程序記憶體中，重新啟動即清空。
07相關的 DSH 工作流程
modlens
作者 liustack
The first vision plugin for DeepSeek Harness, and the vision bridge for every text-only coding agent. Paste an image, get structured JSON evidence (OCR, layout, semantics). | 全網最強 DeepSeek Harness 外掛視覺外掛，為 DeepSeek、GLM 等純文本模型外掛視覺能力，貼上圖片即得結構化 JSON 證據（OCR、版面、語義）。
dsh-vision-router
作者 ysr666
為純文本 DeepSeek Harness 智慧體提供「視覺」能力，內建免金鑰視覺鏈路與畫素級視覺工具，一條命令安裝，無需 Python。
dsh-vision-toolkit
作者 anionex
[dsh]為純文本模型設計更強大的視覺工具箱：一行安裝使用、貼上圖片直接識別、多張圖片問答、截圖到前端UI 還原等｜DeepSeek Harness-native integration for agent-vision-toolkit: image Q&A, long-screenshot OCR, UI restoration, grounding, pixel diff, Artifacts, and Web UI.
dsh-image-gen
作者 shanliuling
直接在 DeepSeek Harness 對話中生成圖片，無需切換外部工具。
08資料與來源
The new [`dsh-vision-toolkit`](https://github.com/Anionex/dsh-vision-toolkit) linked package brings this toolkit into DS…
It provides 10 structured visual tools for intent-aware image Q&A, grounding, detection, tracing, cropping, pixel diff,…
The package is tracked here as a Git submodule and maintained independently at [`Anionex/dsh-vision-toolkit`](https://gi…
此頁面根據專案公開文件、儲存庫中繼資料與 DSH Plugins 的結構化解析所產生；最後核實於 2026-08-21。發現錯誤？提交更正。
最佳 DeepSeek Harness 外掛
從全目錄挑出的 12 個值得優先安裝的外掛，涵蓋各個分類。