agent-vision-toolkit starts from an argument rather than a model: an agent's vision capability doesn't have to live in the model — it can live in the harness. The project has gathered 1,005 stars and 36 forks in the 17 days since its 1 August 2026 creation, one of the faster climbs in the agent-tooling space this month. The target user is someone running a coding agent on a strong but text-only model — DeepSeek is the running example — who keeps hitting a wall the moment an image enters the conversation.

The toolkit ships two layers that can be adopted independently. The first is a set of vision tool CLIs plus a vision-tools skill that teaches the agent when to reach for each one; any agent that can run a shell command can use them. The second, optional layer is seamless integration: a transparent local proxy and single-file native plugins, so that pasted images and the agent's own built-in image tools both start working without extra prompting. The proxy path covers Codex and Claude Code, while Pi, Oh My Pi and OpenCode use native single-file extensions instead.

The individual tools are narrow by design. glance answers a question about an image or transcribes its text, and the key claim is that it forwards the model's current intent rather than returning a generic caption — you get the detail the turn actually needs. ground returns a bounding box in original-image pixel coordinates, with a --region flag that re-searches a sub-box and still reports original coordinates, which is the zoom-in path for small UI targets. detect, crop and trace fill out object listing, cropping and vector extraction. Dependencies stay proportional: glance needs only Python 3.11+, ground/detect/crop add pillow, and trace adds numpy (plus vtracer only for its explicit --outline fallback).

What lifts this above a wrapper collection is the playbook library inside the skill. Documented workflows cover long-screenshot and chat-history OCR — finding low-content cut bands, OCRing chunks in order, preserving speakers and timestamps, merging only true overlap and flagging risky boundaries — plus rebuilding a UI from a screenshot, restoring an icon or logo as a transparent PNG or editable SVG, converting a sketch or whiteboard diagram into Mermaid or Graphviz, and operating a GUI one verified action at a time. Each playbook specifies tool order and how to verify the result, which is the part most such toolkits leave to the model.

Configuration is three environment variables — VISION_API_KEY, VISION_BASE_URL, VISION_MODEL — against any endpoint speaking OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages, so OpenRouter, Aliyun DashScope and similar all work. The most recent addition, on 13 August 2026, is native DeepSeek Harness support through the linked dsh-vision-toolkit submodule, which exposes 10 structured visual tools including grounding, tracing, pixel diff, foreground extraction, dominant-color analysis and HTML screenshots.

Two caveats are worth stating. The toolkit is a router, not a vision model — quality is bounded by whichever multimodal API you point it at, and you pay for those calls. And the default output language is Chinese; English descriptions require setting LANG=en, which is easy to miss. The project reports end-to-end verification in real Codex + DeepSeek sessions and live runs in Claude Code, Pi, Oh My Pi and OpenCode. Licensed MIT.