Opencode Vision

OpenCode Vision

**Disclaimer:** OpenCode Vision is an independent, community-built project. It is **not** built by, endorsed by, or affiliated with the OpenCode team. "OpenCode" appears in this project's name solely to indicate that it builds upon the OpenCode platform.

Introduction

Give text-only OpenCode orchestrators (GLM 5.2, DeepSeek, and similar models) eyes by delegating visual tasks to dynamically registered vision subagents.

Installation

opencode plugin opencode-vision -g

OpenCode installs the npm package via Bun on next launch. Restart OpenCode for the change to take effect.

Configure at least one provider with an image-capable model (`enabled_providers` and/or `provider` entries in OpenCode config). The plugin discovers models from your configured providers and OpenCode's cached model catalog — it does not ship a fixed model list.

Usage

The plugin ships a `vision` skill. When your orchestrator model is text-only and a task needs pixels — not just accessibility metadata — the skill routes the work to a `vision-*` subagent backed by a vision-capable model you choose.

1. Running Visual Tasks

**Image as User Inputs:**

Drag an image into the OpenCode input box, or reference a screenshot path in your message.



**Image as Tool Results:**

The same flow applies to screenshots from browser-use and computer-use tools (chrome-devtools, Playwright, cua-driver, and similar).


2. Picking the Vision Model

On the first visual task, the orchestrator runs the bundled model discovery script and presents a short list of image-capable models from your configured providers. Pick one — that selection is persisted for future sessions.

![Discovering vision-capable models](docs/images/discovering-vision-