pi-qwencloud-provider 🚀

QwenCloud provider for pi — Qwen3.7, DeepSeek V4, GLM-5.2, Wan image generation, and HappyHorse video generation through QwenCloud's OpenAI-compatible API.

QwenCloud uses a static API key (no OAuth) and exposes an OpenAI-compatible Chat Completions API, so no custom streaming protocol is needed — pi's built-in openai-completions streaming handles SSE parsing, tool calls, and usage tracking.

📦 Installation

As a pi extension (recommended)

pi install npm:pi-qwencloud-provider
# or from git
pi install git:github.com/jellydn/pi-qwencloud-provider
# or local path
pi install /path/to/pi-qwencloud-provider

As an npm package

npm install pi-qwencloud-provider
# or
pnpm add pi-qwencloud-provider

Note: This package requires @earendil-works/pi-ai and @earendil-works/pi-coding-agent as peer dependencies. They are automatically available when installed as a pi extension; install them manually when using as a standalone npm dependency.

Pre-requirements

• pi coding agent
• A QwenCloud Token Plan subscription

Features

• Full streaming via QwenCloud's OpenAI-compatible /chat/completions endpoint — SSE parsing, tool calls, and usage tracking handled by pi's built-in openai-completions streaming
• Per-model thinking level support — maps pi's 6 thinking levels (off / minimal / low / medium / high / xhigh) to provider-specific reasoning_effort values, with per-model capability matrices (see table below)
• Per-token cost tracking against QwenCloud reference pricing
• Static API key auth — no OAuth flow; paste your key once via pi /login or set QWENCLOUD_API_KEY
• API key auto-discovery from the QWENCLOUD_API_KEY env var or ~/.pi/agent/auth.json
• Dynamic model discovery — fetches the live model list from the QwenCloud /models endpoint at startup (5s timeout), falling back to a curated static list on error
• /login integration — opens the QwenCloud dashboard and prompts for a paste; keys are sanitized (terminal paste wrappers and control chars stripped)
• Modular architecture — 13 focused source modules (env, auth, thinking, catalog, discovery, models, oauth, error-handler, errors, utils, wan, happyhorse) + entry point (index), all covered by per-module unit tests

Supported Models

Model	Model ID	Context	Reasoning
Qwen3.8 Max Preview	qw/qwen3.8-max-preview	262K	low / medium / high
Qwen3.7 Max	qw/qwen3.7-max	262K	low / medium / high
Qwen3.7 Plus	qw/qwen3.7-plus	1M	low / medium / high
Qwen3.6 Flash	qw/qwen3.6-flash	131K	low / medium / high
DeepSeek V4 Pro	qw/deepseek-v4-pro	1M	high (xhigh → max)
GLM-5.2	qw/glm-5.2	200K	low / medium / high / xhigh

Thinking levels: pi supports 6 levels — off, minimal, low, medium, high, xhigh. Each model declares which levels it supports, mapped to the provider's reasoning_effort parameter. Set the thinking level with pi's --thinking flag or /thinking command. A level marked as unsupported maps to null — no reasoning_effort is sent to the API, so the model runs with its default reasoning behavior.

Image & video generation: Wan (image) and HappyHorse (video) are available via slash commands — /wan for image generation and /happyhorse for video generation. They use separate API endpoints (not chat/completions) and are not available for chat.

Authentication

QwenCloud uses static API keys (no OAuth). The extension resolves the key in this order:

1. Explicit key passed to pi /login
2. QWENCLOUD_API_KEY environment variable
3. ~/.pi/agent/auth.json — { "apiKey": "..." }, { "qw": "..." }, or { "qw": { "access": "..." } }

Set the environment variable:

echo 'export QWENCLOUD_API_KEY="your_key_here"' >> ~/.zshrc
source ~/.zshrc

Or run pi /login and select QwenCloud — it opens the dashboard and prompts you to paste your key (sanitized on input).

Usage

# Non-interactive
pi --model qw/qwen3.7-plus -p "Explain async/await in JavaScript"

# Interactive
pi --model qw/deepseek-v4-pro

# List available models
pi --list-models qw

# Generate an image with Wan
pi --model qw/qwen3.7-plus -p "/wan a cyberpunk cat on a rainy street"

# Generate a video with HappyHorse
pi --model qw/qwen3.7-plus -p "/happyhorse a sunset over the ocean"

# Use in another project
cd my-project
pi --model qw/glm-5.2 --trust "Refactor the auth module"

Switch models in-session with /model qw/qwen3.7-max.

Thinking levels

Set the reasoning effort per model using pi's --thinking flag or the in-session /thinking command:

# Use high reasoning with DeepSeek V4 Pro
pi --model qw/deepseek-v4-pro --thinking high -p "Design a scalable microservice architecture"

# GLM-5.2 supports four levels up to xhigh
pi --model qw/glm-5.2 --thinking xhigh -p "Solve this complex math proof"

# Disable reasoning for a quick code gen task
pi --model qw/qwen3.6-flash --thinking off -p "Write a React form component"

Each model's supported thinking levels are listed in the Supported Models table above. Unsupported levels are not sent to the API — the model runs with its default reasoning behavior.

Environment Variables

Variable	Description	Default
QWENCLOUD_API_KEY	Your QwenCloud API key	—
QWENCLOUD_API_BASE	Override the API base URL	https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1

Run tests

npm test

Pre-commit

This project uses prek to enforce code quality. To install hooks:

prek install

Notes

• Pricing: Per-token costs are reference values for usage tracking only. QwenCloud uses tiered pricing by input length for long-context models — the values in the catalog are entry-tier estimates and may not reflect your actual bill.
• Context windows: taken from QwenCloud model docs — verify against the live /models endpoint.
• Custom API base: set QWENCLOUD_API_BASE env var to override the endpoint.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📋 Changelog

See CHANGELOG.md for release history.

author

👤 Huynh Duc Dung

• Website: https://productsway.com/
• Twitter: @jellydn
• Github: @jellydn

Show your support

Give a ⭐️ if this project helped you!