oc-crofai

An OpenCode plugin that adds CrofAI as a first-class provider.

CrofAI provides cheap access to open-weight LLMs. This plugin dynamically discovers available models from CrofAI's API, so you always see the latest models without updating the plugin.

Installation

opencode plugin -g oc-crofai

Removal

# 1. Remove from OpenCode's config
vi ~/.config/opencode/opencode.json
# Remove the oc-crofai entry from the plugins array

# 2. Delete the plugin package
rm -fr ~/.cache/opencode/packages/oc-crofai@latest/

# 3. Delete the model cache
rm ~/.cache/opencode/crofai-models.json

Usage

1. Get an API key from crof.ai
2. OpenCode will prompt you to enter your CrofAI API key when you first use the provider
3. Select "CrofAI" as your provider and choose a model

Alternatively, set the CROFAI_API_KEY environment variable:

export CROFAI_API_KEY="your-api-key-here"

How It Works

The plugin uses three OpenCode hooks:

• config — Registers CrofAI as a provider with @ai-sdk/openai-compatible SDK integration
• provider — Fetches https://crof.ai/v1/models on startup and maps the response (context length, pricing, reasoning capabilities) to OpenCode's model schema. Models are sorted alphabetically and cached to disk for instant startup, with background refresh to keep them current.
• auth — Provides interactive API key entry with automatic CROFAI_API_KEY env var fallback

Features

• Disk cache with background refresh — Models load instantly from cache on startup; fresh data is fetched in the background so subsequent calls always see the latest models
• 1-hour cache TTL — Stale caches are discarded and re-fetched automatically
• Cache schema version — Old cached data from previous plugin versions is automatically discarded
• 10s fetch timeout — Prevents hangs if the API is unreachable
• Model sorting — Models are sorted alphabetically for consistent ordering in the UI
• Vision detection — Vision-capable models are automatically detected by scraping CrofAI's pricing page; no hardcoded list to maintain
• Reasoning variants — All reasoning-capable models include variants for controlling thinking depth (none/low/medium/high)

Vision Support

Vision-capable models are automatically detected from CrofAI's pricing page, so models with image input are marked accordingly in the model list.

Thinking / Reasoning Support

Reasoning variants are injected via the config hook, based on teppyboy/opencode-crofai's approach.

All reasoning-capable models include four variants for controlling thinking depth:

Variant	Reasoning Effort	Effect
none	"none"	Disables reasoning — reduces latency and cost
low	"low"	Minimal thinking, faster responses
medium	"medium"	Balanced (default)
high	"high"	Maximum thinking, best reasoning

To use a variant, set it in your OpenCode agent config:

agent:
  build:
    model: crofai/kimi-k2.6
    variant: high

License

MIT