modlens

Give a text-only model sight: paste an image straight into the chat and it reads like it can see.

What problem it solves

The flagship DeepSeek and GLM chat models are text-only and cannot read images. To “read” a screenshot, an error, or a chart mid-conversation, you usually switch to a multimodal model or save the image to a file and pass a path. ModLens makes it a straight paste: a plug-in vision engine that gives a text-only model sight, with no file-saving and no separate vision model or API key to configure.

It exposes a native modlens_read_image tool and auto-discovers every provider route carrying text-only DeepSeek or GLM models, adding a wrapped (modlens vision) entry per route (it remembers your pick, so once is enough). There are two paste routes: just paste (the image lands as a private temp file), or pick the (modlens vision) entry and paste (the thumbnail stays visible in your message, closer to the Codex feel); images are converted to structured evidence at request time and answered by the same underlying route.

Key features

• Paste-to-read: paste an image and it reads it — no saving to a file and passing a path first
• Native tool and routes: exposes modlens_read_image and adds a (modlens vision) entry per text-only DeepSeek / GLM route, automatically excluding vision models
• Six built-in engines + failover: gemini-api, openai, anthropic, antigravity-cli, claude-cli, kimi-cli form one failover chain; meta.attempts records every try so fallback is never silent
• Zero-config start: reuses the multimodal logins Claude Code, Codex, OpenCode, or Pi already have; with nothing set up, Antigravity CLI is a free no-key channel
• Evidence, not imagination: full transcription, reading-order layout regions, and entity / relation lists — the model quotes specifics
• openai as a universal socket: any OpenAI-compatible endpoint (qwen-vl, GLM, SiliconFlow, OpenRouter, self-hosted vLLM / Ollama) plugs straight in

Quick start

One command to install (the version is named explicitly so pnpm 11 resolves it correctly):

npx -y @deepseek-ai/dsh plugin —profile web add @liustack/modlens@3.22.0

After that, just chat: paste an image or drop a path and the skill triggers on its own — the image goes to a vision engine and the answer comes back grounded in what it read. Configure engines with modlens config set provider <name> (preference) or -p <name> (pin exactly one); behind a proxy set modlens config set proxy <url>.

Who it is for

• DSH users on text-only models who often need to read images (screenshots, errors, charts)
• Developers who don’t want to set up a separate vision model or API key just for image reading
• Users who want verifiable image answers that cite specific evidence rather than vague descriptions