See

An OpenCode plugin that adds an opencode_see tool. It sends an image to a vision-capable LLM and returns a text description, falling back across three free providers — Gemini, Groq, and Cerebras — one at a time, in whatever order you configure.

No merging of results: it tries provider #1, and only moves on to #2 if #1 fails or isn't configured. Whichever one succeeds first wins.

Features

• One opencode_see tool — point it at one or more local files or http(s) URLs, sent in a single request.
• Free providers: Gemini, Groq, and Cerebras — no API cost.
• One-at-a-time fallback: tries each provider (and each of its models) in order until one succeeds; results are never merged.
• Configurable provider order and per-provider model fallback lists.

1. Get free API keys

Provider	Where	Env var
Gemini	aistudio.google.com	GEMINI_API_KEY
Groq	console.groq.com	GROQ_API_KEY
Cerebras	cloud.cerebras.ai	CEREBRAS_API_KEY

You don't need all three — any subset works. Providers with no key set are skipped automatically.

2. Install

As a project or global plugin (from npm):

npm install opencode-see

Then add it to opencode.json:

{
  "plugin": ["opencode-see"]
}

Or as a local plugin, copy dist/plugin.js into .opencode/plugins/opencode-see.js (project-level) or ~/.config/opencode/plugins/opencode-see.js (global).

3. Set your keys

You can provide keys (and other settings) either via plugin options in opencode.json or via environment variables. Env vars are used as the fallback whenever an option isn't set.

Option A — opencode.json plugin options. Register the plugin in tuple form and pass apiKeys (and optionally models, order, prompt):

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    ["opencode-see", {
      "apiKeys": {
        "gemini": "...",
        "groq": "...",
        "cerebras": "..."
      },
      "models": {
        "gemini": ["gemini-3-flash-preview", "gemini-2.5-flash", "gemini-3.1-flash-lite"],
        "groq": ["qwen/qwen3.6-27b"],
        "cerebras": ["gemma-4-31b"]
      },
      "providerOrder": "cerebras,gemini,groq",
      "defaultPrompt": "Describe this image in detail."
    }]
  ]
}

Option B — environment variables. Export whichever keys you have, e.g. in your shell profile or .env (however you normally get env vars into the OpenCode process):

export GEMINI_API_KEY="..."
export GROQ_API_KEY="..."
export CEREBRAS_API_KEY="..."

Either way, you don't need all three — any subset works, and providers with no key set are skipped automatically.

Precedence for every setting (highest wins): tool argument → opencode.json plugin option → environment variable → built-in default.

Providers and models: for each provider you can list multiple models in fallback order via the models array — the plugin tries each one until one succeeds, then moves to the next provider. A provider with no models configured uses its env var (or built-in default).

4. Configure provider order (optional)

Default order is gemini,groq,cerebras. Override globally via the providerOrder plugin option (above) or the env var:

export OPENCODE_SEE_PROVIDER_ORDER="cerebras,gemini,groq"

Or override per-call by passing providers as a tool argument (the agent can do this itself if you ask, e.g. "describe this image, try cerebras first").

Other settings:

• defaultPrompt plugin option or OPENCODE_SEE_DEFAULT_PROMPT env var — default prompt when none is given.
• models plugin option (array of model ids, tried in order) or GEMINI_VISION_MODEL, GROQ_VISION_MODEL, CEREBRAS_VISION_MODEL env vars — override the model id per provider (useful since Groq in particular rotates its vision model lineup). Env vars accept a single model; the models option accepts several, tried one at a time.

5. Use it

Just ask the OpenCode agent to look at an image:

"Describe screenshot.png" "What does the error in ./logs/crash.jpg say?" "Look at https://example.com/diagram.png and explain the architecture" "Compare these two screenshots: after/login.png and after/home.png"

The image argument is an array, so the agent passes "image": ["a.png", "b.png"] for multiple images (a single image is a one-element array).

The tool result shows which vision model answered — the output starts with a **Vision model:** <Provider> (<model>) line followed by the model's description (e.g. **Vision model:** Gemini (gemini-3.1-flash-lite)).

Known limits

• Cerebras caps images at 280 tokens — fine for simple screenshots, weak for dense/detailed images. Best used as a fallback, not primary.
• Groq's vision model id changes periodically as they retire/replace models. If it stops working, check https://console.groq.com/docs/vision and set GROQ_VISION_MODEL (or add it to the models option).
• No client-side image resizing yet (large images are sent as-is). If you hit payload-size or token-limit errors on big screenshots, downscale before passing the path, or open an issue — this is a natural next addition.

Development

npm install
npm run build

Note: npm test is currently stale/broken — the script points at dist/test/*.test.js, but no test directory exists yet. Don't rely on it.

Source is in src/, one file per provider under src/providers/, plus:

• image.ts — loads local files / URLs into a base64 data URI
• config.ts — resolves provider order from args/env/defaults
• orchestrator.ts — the one-at-a-time fallback loop
• plugin.ts — registers the opencode_see tool with OpenCode

Publishing

dist/ is gitignored but shipped to npm via the files field. Publish with:

npm run build        # build (also runs automatically via prepublishOnly)
npm publish          # pack and publish

Use npm publish --dry-run first to preview exactly what gets packed. Bump the version with npm version patch|minor|major before each release.

License

MIT