modlens: giving the model eyes

The first vision plugin: paste an image, a vision engine returns structured JSON evidence, and the model answers from the evidence.

Config fragment

bash

# pin the version - pnpm intentionally delays fresh publishes for 24h
npx -y @deepseek-ai/dsh plugin --profile web add @liustack/modlens@3.18.1

The plugin keeps your API keys self-hosted: Gemini, OpenAI-compatible or Claude CLI engines are configured locally, never sent to a third-party proxy.

Core code

The three-step data flow is the teardown highlight: image in -> structured evidence out -> model answers.

// 1. paste image -> 2. vision engine returns JSON evidence
const evidence = await visionEngine.analyze(image, {
  ocr: true,
  layout: true,
  entities: true,
});
// 3. evidence becomes the tool result the model reasons over
return { ok: true, evidence };

OCR transcription, reading-order layout and entity relations are returned as structured JSON, not prose — exactly what a reasoning model can consume.

Why it scores

First-mover: the first vision plugin in the ecosystem (857 to ~2444 stars).
Design: external capability returning structured evidence is the canonical tool pattern.
Privacy: keys stay on the machine; nothing is proxied.