MIT 许可的纯文本模型视觉桥接插件，用一条 `dsh plugin add` 命令装进 DeepSeek Harness（dsh）——粘贴图片即得结构化 JSON 证据（OCR、版面、语义）。
DSH 适配
兼容
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
MIT
01它能帮你完成什么？
Read images pasted straight into DeepSeek Harness chats running text-only DeepSeek / GLM models
Grounded answers built on structured JSON evidence — full transcription, reading-order layout regions, entity and relation lists
Developers running text-only DeepSeek or GLM models in dsh who need screenshots, slides, documents and charts understood in-conversation
Analyze dense charts, tweet screenshots and multi-image batches without leaving the dsh conversation
Detail-level readings — axes, log scale, color coding, timestamps, engagement numbers, per-image content and style — delivered one by one
dsh users who paste analytics charts, social screenshots or batches of images and need every element quoted back, not summarized away
Keep native paste thumbnails while adding vision via a wrapped model entry
A `(modlens vision)` model-selector entry per text-only route (choice is remembered); pasted thumbnails stay visible and are converted to structured evidence at request time
dsh users who want a Codex-app-like paste experience — visible thumbnails — instead of temp-file paths in the composer
02如何接入 DeepSeek Harness？
前置条件
- DeepSeek Harness (dsh) available — the install runs through the official `npx -y @deepseek-ai/dsh` CLI
- At least one vision engine: a free Gemini API key, the free Antigravity CLI (no key), an Anthropic key, any OpenAI-compatible endpoint, or an existing Claude Code / Kimi / Codex / OpenCode / Pi login reused with consent
安装步骤
- 01
Install the plugin with one command: `npx -y @deepseek-ai/dsh plugin --profile web add @liustack/modlens@3.22.1`
$ npx -y @deepseek-ai/dsh plugin --profile web add @liustack/modlens@3.22.1
- 02
Pick a `(modlens vision)` entry in the model selector (it remembers your choice, so once is enough), or just paste an image on a confirmed text-only model and let `modlens_read_image` take over
- 03
If no vision engine is configured yet, set up a free Gemini API key (about three minutes, no credit card) or install Antigravity CLI and sign in with `agy`
验证接入成功
作者未说明
回滚
- Uninstalling is deleting a folder — no harness config lines are changed, so your agents are back to stock
03DSH 适配与能力边界
Installed inside DeepSeek Harness as exactly one dsh plugin via a single `dsh plugin add` command; adds `(modlens vision)` model-selector entries and reads pasted images through the `modlens_read_image` tool
Direct-paste vision for text-only models
an image pasted straight into the chat, or a dropped file path→structured JSON evidence — full transcription, reading-order layout regions, entity and relation lists — grounding the model's answer
images are sent over the network to an external vision engineon a text-only model a pasted image lands as a private temp file whose path enters the composerTen-engine failover chain
six built-in providers (Gemini, OpenAI-compatible, Anthropic, Antigravity CLI, Claude CLI, Kimi CLI) plus up to four reused local agent CLIs→the first good result wins; `meta.attempts` records every attempt so a fallback is never silent
reused CLI reads spend that CLI owner's quota and are labeled in `meta.warnings`reuse is only enabled after explicit per-harness consent recorded in configUniversal OpenAI-compatible socket
`openai.baseUrl` + `openai.apiKey` + `openai.model` config (e.g. qwen-vl, GLM, SiliconFlow, OpenRouter, self-hosted vLLM/Ollama)→any endpoint speaking the OpenAI chat-completions protocol with image input driven as a vision engine
writes provider keys and endpoints into the modlens config
04适合谁？何时不该用？
适合
- Developers running text-only DeepSeek or GLM models in dsh who need screenshots, slides, documents and charts understood in-conversation
- dsh users who paste analytics charts, social screenshots or batches of images and need every element quoted back, not summarized away
- dsh users who want a Codex-app-like paste experience — visible thumbnails — instead of temp-file paths in the composer
不适合
- ModLens ships no local vision model: every read is routed to an external engine (Gemini / OpenAI-compatible / Anthropic API, Antigravity CLI, or a reused agent CLI), so it needs network access and your use of upstream engines is governed by their own terms and quotas.
05兼容性、维护与安全提示
- ModLens ships no local vision model: every read is routed to an external engine (Gemini / OpenAI-compatible / Anthropic API, Antigravity CLI, or a reused agent CLI), so it needs network access and your use of upstream engines is governed by their own terms and quotas.
- Read speed depends on the engine: API providers take 5-10 seconds per read, while CLI-based engines (Antigravity, Claude CLI, Kimi CLI, reused Codex / OpenCode) take 20-45 seconds.
- Only a model whose metadata positively confirms it is text-only is taken over with a `(modlens vision)` entry; anything unconfirmed is left alone, and the DeepSeek / GLM families' own vision models keep their native paste.
MIT · actively maintained (latest release v3.22.1, 2026-08-20)
06常见问题
如何把 ModLens 安装到 DeepSeek Harness？
运行一条命令：`npx -y @deepseek-ai/dsh plugin --profile web add @liustack/modlens@3.22.1`。插件会自动发现所有承载纯文本 DeepSeek/GLM 模型的 provider 路由，并为每条路由添加一个 `(modlens vision)` 包装入口（默认安装会得到 DeepSeek-V4-Flash 和 DeepSeek-V4-Pro 两个变体）。在模型选择器里选一次即可（选择会被记住），之后直接粘贴图片就行。
它是 dsh 原生插件还是 MCP？
它以 dsh 插件形式安装，在 dsh 里就是一个插件：没有 hooks、没有 wrapper、没有本地代理守护进程，任何 harness 配置文件一行都不改。卸载就是删掉一个文件夹，agent 立即恢复原状。
需要什么视觉引擎或前置条件？
任选一个引擎即可：免费 Gemini API key（约三分钟、无需信用卡，推荐默认，单次读图 5-10 秒）、任意 OpenAI 兼容接口（qwen-vl、GLM、自建网关）、Anthropic key、免 key 的 Antigravity CLI，或者经逐个 harness 明确同意后复用 Claude Code / Kimi / Codex / OpenCode / Pi 的现有登录。
我粘贴的图片会发送到哪里？
图片会通过网络发送给你配置的故障转移链中的视觉引擎；每次尝试都记录在 `meta.attempts` 里，任何消耗了复用 CLI 配额的读取都会在 `meta.warnings` 中标注花了谁的配额。上游引擎各自的条款与配额由你自行负责。
哪些模型会被接管加上视觉？
只有元数据明确确认是纯文本的模型才会被加上 `(modlens vision)` 入口；无法确认的模型保持原样，DeepSeek/GLM 自家的视觉模型也保留原生粘贴行为。
07相关的 DSH 工作流
agent-vision-toolkit
作者 anionex
为纯文本模型"看图“设计更好的视觉工具箱和技能，支持多图理解，图片问答，前端UI还原、GUI 自动化等，并可选无缝接入多个主流agent，直接识别粘贴图片｜ A vision toolkit and skill designed for text-only llms — image Q&A, long-screenshot OCR, frontend UI restoration, and GUI automation, with optional seamless integration for Codex, Claude Code, Pi, Oh My Pi, and OpenCode
dsh-vision-router
作者 ysr666
为纯文本 DeepSeek Harness 智能体提供「视觉」能力，内置免密钥视觉链路与像素级视觉工具，一条命令安装，无需 Python。
dsh-vision-toolkit
作者 anionex
[dsh]为纯文本模型设计更强大的视觉工具箱：一行安装使用、粘贴图片直接识别、多张图片问答、截图到前端UI 还原等｜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 对话中生成图片，无需切换外部工具。
08数据与来源
**🥇 The most capable vision plugin for DeepSeek Harness (dsh):** install it instantly with one command: `npx -y @deepse…
on the skill harnesses it is exactly one skill folder, on dsh exactly one plugin. Uninstalling is deleting a folder, and…
The plugin auto-discovers every provider route carrying text-only DeepSeek or GLM models and adds a wrapped entry per ro…
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-21。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。