openqwencode

Local OpenCode plugin for qwen.ai OAuth, with the same persistent credential cache style as qwen-code and opencode-qwen-auth, but without the legacy built-in Qwen providers/models.

This plugin builds on the newer Qwen Code direction and refresh work, adapted for OpenCode as a dedicated plugin-based integration.

It was also created because older OpenCode Qwen auth plugins like 1579364808/opencode-qwen-auth and gustavodiasdev/opencode-qwencode-auth no longer work.

Reviewed external references

This plugin was built after reviewing the main public Qwen/OpenCode auth implementations and forks, not just a single upstream:

• QwenLM/qwen-code: primary upstream reference for OAuth flow, credential persistence, and refresh behavior
• foxswat/opencode-qwen-auth: strongest reference for resilience improvements and OpenCode integration ideas
• RunMintOn/OpenCode-Qwen-Proxy: useful reference for throttling and burst-handling behavior
• 1579364808/opencode-qwen-auth: reviewed as an older baseline
• gustavodiasdev/opencode-qwencode-auth: reviewed as another older OpenCode-oriented implementation

The current design intentionally follows the strongest parts of those implementations while keeping the plugin small, current, and focused on the official free Qwen OAuth path.

Language rule

• Everything in this project must be written in English.

What this plugin does

• login provider: openqwencode
• model path: openqwencode/coder-model
• only the official free OAuth model: coder-model
• image input through the same coder-model
• tokens stored in qwen-code-compatible format in ~/.qwen/oauth_creds.json
• automatic refresh so you do not have to log in again every day
• atomic writes + lockfile so multiple processes do not corrupt credentials
• in-memory cache with periodic disk sync
• backoff for 429 and temporary server errors
• cancellable auth polling, including process-signal cancellation
• removes/overrides legacy qwen and qwen-code providers

OpenCode config

{
  "$schema": "https://opencode.ai/config.json",
  "disabled_providers": ["qwen", "qwen-code"],
  "plugin": ["file:///ABSOLUTE/PATH/TO/openqwencode"]
}

Login:

opencode auth login -p openqwencode

Then use:

opencode --model openqwencode/coder-model

Install and use in OpenCode

Install the package:

npm install openqwencode

Add it to your OpenCode config:

{
  "$schema": "https://opencode.ai/config.json",
  "disabled_providers": ["qwen", "qwen-code"],
  "plugin": ["openqwencode"]
}

Authenticate:

opencode auth login -p openqwencode

Run OpenCode with the plugin model:

opencode --model openqwencode/coder-model

Link this install to your telemetry dashboard connection:

npx openqwencode link

Optional environment variables

Keep these unset unless you explicitly want the behavior.

• OPENQWENCODE_BASE_URL: overrides the upstream OpenAI-compatible API base URL. If set to a bare origin such as https://qwen.2631.eu, the plugin normalizes it to /v1 automatically.
• OPENQWENCODE_TELEMETRY_ENDPOINT: sends a best-effort telemetry POST after each final model request response.
• OPENQWENCODE_TELEMETRY_LINK_URL: optional override for the device-link service base URL. If unset, the plugin derives it from OPENQWENCODE_TELEMETRY_ENDPOINT.

Example:

export OPENQWENCODE_BASE_URL="https://qwen.2631.eu"
export OPENQWENCODE_TELEMETRY_ENDPOINT="https://qwen.2631.eu/telemetry"
# optional if your link flow lives elsewhere
export OPENQWENCODE_TELEMETRY_LINK_URL="https://qwen.2631.eu"

The telemetry payload is intentionally minimal:

{
  "ts": "2026-04-13T12:34:56.789Z",
  "deviceCode": "a1b2c3d4e5f6",
  "requestCount": 1,
  "attempts": 1,
  "status": 200,
  "durationMs": 842,
  "path": "/v1/chat/completions"
}

Notes:

• your telemetry server will already see the client IP at the HTTP layer, so the plugin does not send it explicitly
• deviceCode is a deterministic short hash derived from the local machine identity so you can distinguish computers without sending raw host details
• the device-link flow stores its long-lived telemetry token separately in ~/.qwen/telemetry_creds.json
• run npx openqwencode link to open the dashboard link page, enter the short code shown in the terminal, and attach this computer to one shared connection bucket
• telemetry is fire-and-forget, best-effort, and disabled by default
• no official public Qwen Code reset time could be confirmed; use the UTC timestamp (ts) to inspect your own reset pattern from collected telemetry

If you upgraded from an older published version and OpenCode still does not see the model, clear the cached package and refresh:

rm -rf ~/.cache/opencode/packages/openqwencode@latest ~/.cache/opencode/node_modules/openqwencode
opencode models --refresh

Notes

• upstream QwenLM/qwen-code only uses coder-model for the free OAuth flow
• qwen3-coder-plus, qwen3-vl-plus, and vision-model are intentionally no longer exposed
• the plugin injects the Qwen OAuth headers and a minimal Qwen Code system message
• the local credential file is the primary source of truth; OpenCode auth state is only used as a bootstrap/migration path
• project rules, release workflow, and backlog now live in AGENTS.md