Skill Detail

Give text-only agents vision workflows with Agent Vision Toolkit

Install Agent Vision Toolkit so shell-capable coding agents can inspect screenshots, run OCR, locate UI elements, restore interfaces, and operate GUIs through repeatable local vision playbooks.

Image & Creative AutomationMulti-Framework
Image & Creative Automation Multi-Framework Security Reviewed
⭐ 335 GitHub stars
INSTALL WITH ANY AGENT
npx skills add agentskillexchange/skills --skill give-text-only-agents-vision-workflows-with-agent-vision-toolkit Copy
Works best when you want a reusable capability, not another fragile one-off prompt.
At a glance
Tools required
Python 3.11+, shell access, an OpenAI-compatible vision API for model-backed image tools, optional Pillow/numpy/vtracer for advanced local operations
Install & setup
Configure VISION_API_KEY, VISION_BASE_URL, and VISION_MODEL in ~/.config/agent-vision-toolkit/env, then restrict that file to the current user.

git clone https://github.com/Anionex/agent-vision-toolkit.git
export PATH="$PWD/agent-vision-toolkit/bin:$PATH"

Copy agent-vision-toolkit/skills/vision-tools/ into your agent skills directory and restart the agent. Review the upstream AGENT_INSTALL.md only when adding the optional proxy or native integration for your runtime.
Author
Anionex
Publisher
Independent
Last updated
Aug 7, 2026
Quick brief

Agent Vision Toolkit gives text-only agent runtimes a repeatable way to handle visual work without switching the whole session to a multimodal model. The upstream project ships local CLIs for image description and OCR, object grounding, UI element detection, crop extraction, and SVG tracing, plus a vision-tools Agent Skill that tells the agent when to use each tool and how to chain them.

How it works

What this skill actually does

Use this when an operator wants an agent to answer questions about screenshots, split and OCR long scrolling captures, locate controls before GUI actions, rebuild UI from a reference image, restore graphics, or verify visual changes with pixel-aware checks. It is especially useful when the active model is text-only but the harness can invoke shell tools or accept a local proxy or native extension.

The scope boundary is the visual-agent workflow: installing the toolkit, configuring an OpenAI-compatible vision endpoint, exposing commands such as glance, ground, detect, trace, and crop, and following the included playbooks for OCR, UI restoration, graphic restoration, structure recovery, and GUI operation. This is not a generic computer-vision library listing and not just a product card; the skill value is the repeatable operator procedure for text-only agents across Codex, Claude Code, Pi, Oh My Pi, OpenCode, and other shell-capable runtimes.