at-vision

Inspect screenshots, diagrams, and image URLs through a configured vision model.

180|Updated Jul 7, 2026
One-click install
npx skills add https://github.com/kairyou/agent-tools --skill at-vision-kairyou
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: at-vision
Source: https://github.com/kairyou/agent-tools/tree/main/capabilities/vision/skills/at-vision
Command: npx skills add https://github.com/kairyou/agent-tools --skill at-vision-kairyou

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? When a task depends on visible content but the prompt only contains an image file path or URL, the agent cannot see the image. This Skill routes one image plus narrow factual questions to a configured vision model via an MCP tool or installed CLI, returning observations the agent can reason over. ## Core Features & Use Cases - Targeted visual questions: Ask narrow factual questions (error codes, hex colors, pixel spacing) about a local file or image URL and receive per-question answers with uncertainty notes. - Whole-image transcription: Extract structured output from design mockups (HTML skeleton with inline CSS), text-heavy documents (Markdown), or charts (Markdown data tables). - MCP tool with CLI fallback: Prefer the inspect_image MCP tool from the agent-tools-vision server, falling back to the installed CLI invoked with a temporary JSON request file. - Use Case: A user shares a path to a UI mockup screenshot and asks you to implement it. The Skill transcribes the mockup as HTML with hex colors and pixel sizes, then answers follow-up questions to verify details. ## Quick Start Ask the agent to inspect the screenshot at ./designs/homepage.png and transcribe it as an HTML skeleton with inline CSS.

Frequently Asked Questions about at-vision

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I inspect an image file or URL with an AI agent?▼

Provide the image as a file path or http(s) URL along with narrow factual questions. The inspect_image MCP tool sends one image per call to a configured vision model and returns answers keyed by question id, with optional uncertainty notes.

How to convert a UI design mockup screenshot into HTML?▼

Ask one whole-image question requesting a structured transcription: an HTML skeleton with inline CSS, hex color estimates, pixel sizes, and real text content. Treat returned values as visual estimates and verify against design tokens when exactness matters.

When should the vision inspection tool not be called?▼

Do not call it when the prompt already contains actual image content, when the task only involves file operations like renaming or moving, or when the user says not to send the image to a provider. A bare path without a content-dependent task is also not a reason.

What happens if the vision MCP tool is unavailable?▼

Fall back to the installed CLI by writing a temporary JSON request file with the image source and questions, then running node on the installed CLI path with --request-file. If neither is available, install the capability with npx -y @kairyou/agent-tools@latest vision.

Why does the vision tool report a config_error?▼

A config_error means the vision provider is not configured. Set the provider, baseUrl, model, and apiKey in ~/.agent-tools/config.jsonc as described in the agent-tools README, then retry the inspection.