Amanai Provider

An OpenCode plugin that registers the Amanai provider — an OpenAI-compatible API at https://api.amanai.dev/v1 — and discovers the models it serves (currently ~47: GLM, Qwen, Kimi, DeepSeek, MiniMax, Grok, GPT, Claude, and more).

Install

Publish the package once (see Publishing), then install it on any machine:

# From npm
opencode plugin add opencode-amanai-provider

# Or straight from a Git repository
opencode plugin add github:YOUR_GITHUB_USER/opencode-amanai-provider

Restart or reload OpenCode, then select a model with /models, for example amanai/glm-5.3.

For a local checkout instead of a published package, add the directory to plugins in your global opencode.jsonc:

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["/absolute/path/to/opencode-amanai-provider"]
}

Authentication

The plugin reads your API key from the AMANAI_API_KEY environment variable:

# PowerShell
$env:AMANAI_API_KEY = "your-key"
opencode service restart

# bash / zsh
export AMANAI_API_KEY="your-key"
opencode service restart

You can also run /connect, choose Amanai, and paste the key. The plugin registers both an environment method and an interactive key method.

Configuration

Setting	Default	Purpose
AMANAI_API_KEY	–	API key.
AMANAI_BASE_URL	https://api.amanai.dev/v1	Override the endpoint.

How it works

On setup the plugin fetches GET /models and registers every returned model with its context window, output limit, and input modalities. A baked-in inventory is used when discovery is unavailable, and the inventory is refreshed every six hours.

Amanai advertises the reasoning levels it accepts in each model's thinking list. The plugin exposes every level except the default (auto) as a model variant backed by settings.reasoningEffort, so you can pick one with the #variant selector:

opencode run --model amanai/glm-5.3#high "Review this migration plan"

Run /models to see the available efforts for a model. Levels differ per model (low/medium/high/xhigh/max, and none/minimal where supported).

Transient failures are retried instead of ending the turn. A request that hits a rate limit, a server error, or loses its connection mid-stream is retried up to five times with exponential backoff and jitter, mirroring the Command Code provider. Client errors such as a bad request or a rejected credential stay terminal.

Publishing

npm login
npm publish --access public

The package name opencode-amanai-provider is free on the public registry; use a scope such as @yourname/opencode-amanai-provider if you prefer.