opencode-image-parsing

opencode-image-parsing

English | 简体中文

⚠️ This project is built on opencode's skill + subagent mechanism and tested on opencode. The mechanism itself is a general pattern — other agent frameworks with skill/subagent support (e.g. Claude Code) may work in theory, but the config paths, frontmatter fields, and model IDs (e.g. opencode/mimo-v2.5-free) are written for opencode. Adapt them to the target tool when migrating.

Let text-only models (e.g. DeepSeek V4 Flash) "see" images, via a skill + subagent.

When the model cannot parse an image, it automatically delegates the parsing to MiMo-V2.5 (a vision-capable model), then continues answering based on the parsed text. Your chosen text model stays as the main model — only the "look at the image" step goes through MiMo.

How it works

Text-only model (can't see images)
   │  skill "image-parsing" triggers
   ▼
Task invokes subagent mimo-vision (model: MiMo-V2.5)
   │  reads the image file with the Read tool
   ▼
Returns a detailed text description of the image
   ▼
Text-only model answers based on that text

• Local images (drag & drop) → the file path is passed directly to the subagent.
• Clipboard-pasted images → the base64 payload is decoded from the conversation data into a temp file, then passed to the subagent.

One-line install

Windows (PowerShell):

$d = "$HOME\.config\opencode"
New-Item -ItemType Directory -Path "$d\agent", "$d\skills\image-parsing" -Force | Out-Null
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/starstarstars/opencode-image-parsing/main/agent/mimo-vision.md" -OutFile "$d\agent\mimo-vision.md"
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/starstarstars/opencode-image-parsing/main/skills/image-parsing/SKILL.md" -OutFile "$d\skills\image-parsing\SKILL.md"

macOS / Linux (bash/zsh):

mkdir -p ~/.config/opencode/skills/image-parsing ~/.config/opencode/agent
curl -fsSL -o ~/.config/opencode/agent/mimo-vision.md https://raw.githubusercontent.com/starstarstars/opencode-image-parsing/main/agent/mimo-vision.md
curl -fsSL -o ~/.config/opencode/skills/image-parsing/SKILL.md https://raw.githubusercontent.com/starstarstars/opencode-image-parsing/main/skills/image-parsing/SKILL.md

Restart opencode after installing.

Manual install

Copy skills/image-parsing/ and agent/mimo-vision.md to either location (or both):

Global install (all projects):

~/.config/opencode/skills/image-parsing/SKILL.md
~/.config/opencode/agent/mimo-vision.md

Project-level install:

.opencode/skills/image-parsing/SKILL.md
.opencode/agent/mimo-vision.md

Auth (free)

The default subagent uses opencode/mimo-v2.5-free:

opencode auth login

One login gives you a free OpenCode Zen key. The model is free and supports image input — no opencode-go subscription needed.

Switching models

To use a different model, change the model field in agent/mimo-vision.md:

Restart opencode after changing the config.

In action

After pasting a screenshot, the model outputs something like:

Image detected (pasted) → invoking mimo-vision → parsed → continuing...

This is an e-commerce product page: title "xxx", price ¥xx, xx in stock...

Notes

• Internal details (base64 decoding, temp files, conversation-data extraction) are kept out of the output.
• If the subagent's model is not authorized, it will prompt you to run opencode auth login or switch models — it won't get stuck.

License

MIT © starstarstars