ZeroGPU Router

OpenCLAW skill and CLI for routing focused tasks directly to the ZeroGPU API.

The main production path does not require a hosted adapter. The zerogpu-router CLI calls:

https://api.zerogpu.ai/v1/responses

with the user's x-api-key and x-project-id headers.

Quick Start For Users

Install and configure with OpenCLAW's native plugin flow:

openclaw plugins install zerogpu-router
openclaw gateway restart

This registers ZeroGPU as an available provider and exposes task offload tools such as zerogpu_summarize, zerogpu_classify, zerogpu_extract, and zerogpu_followups. It does not change the user's primary model. Keep your normal OpenCLAW model, such as nearai/auto, as the default brain.

Configure credentials with the hosted installer:

curl -fsSL https://raw.githubusercontent.com/zerogpu/ZeroGPU-OpenClaw-Plugin/main/scripts/setup-openclaw-provider.sh | bash

The installer asks for:

• ZeroGPU API key
• ZeroGPU project ID

It stores those credentials locally in ~/.openclaw/zerogpu/config.json, and installs local zerogpu skill guidance so the normal agent knows when to call the offload CLI. Your primary model remains unchanged.

For OpenCLAW Cloud environments where restart is handled by the UI:

curl -fsSL https://raw.githubusercontent.com/zerogpu/ZeroGPU-OpenClaw-Plugin/main/scripts/setup-openclaw-provider.sh | SKIP_GATEWAY_RESTART=1 bash

Then restart or reload the gateway from the OpenCLAW Cloud UI.

Verify the skill and provider are visible:

openclaw config get models.providers.zerogpu
openclaw skills list | grep -i zerogpu
$HOME/.openclaw/bin/zerogpu-router summarize "Summarize this sentence."

To intentionally make ZeroGPU the global default model, opt in explicitly:

curl -fsSL https://raw.githubusercontent.com/zerogpu/ZeroGPU-OpenClaw-Plugin/main/scripts/setup-openclaw-provider.sh | SET_ZEROGPU_AS_DEFAULT=1 bash

Local Plugin Development

If you want to inspect or modify the plugin locally:

git clone https://github.com/zerogpu/ZeroGPU-OpenClaw-Plugin.git
cd ZeroGPU-OpenClaw-Plugin
npm install
npm run check

You can still run the fallback setup helper locally:

ZEROGPU_API_KEY="YOUR_ZEROGPU_API_KEY" \
ZEROGPU_PROJECT_ID="YOUR_ZEROGPU_PROJECT_ID" \
npm run setup:openclaw

Optional Hosted Adapter

The Render adapter is optional/legacy. It is useful only if you want a shared OpenAI-compatible bridge endpoint with dashboard events:

1. Create a Render Web Service from this repo.
2. Use build command npm install.
3. Use start command npm start.
4. Set health check path /health.

Do not add ZEROGPU_API_KEY or ZEROGPU_PROJECT_ID to Render. Users provide those during OpenCLAW setup.

Local Development

npm install
npm start

The server listens on http://localhost:8787 by default.

Useful checks:

curl http://localhost:8787/health
curl http://localhost:8787/v1/models
curl http://localhost:8787/v1/zerogpu/models

Debug calls can pass credentials directly:

curl -X POST http://localhost:8787/v1/chat/completions \
  -H "content-type: application/json" \
  -H "x-api-key: YOUR_ZEROGPU_API_KEY" \
  -H "x-project-id: YOUR_ZEROGPU_PROJECT_ID" \
  -d '{
    "model": "zerogpu/summarize",
    "messages": [
      { "role": "user", "content": "Summarize this feature request in 3 bullets." }
    ]
  }'

The old endpoint remains available as a compatibility alias:

• POST /v1/zerogpu/chat/completions

Model Aliases

• zerogpu/auto routes based on task detection.
• zerogpu/chat uses general chat.
• zerogpu/chat-thinking uses reasoning-oriented chat.
• zerogpu/summarize routes summarization.
• zerogpu/classify routes classification.
• zerogpu/extract routes extraction.
• zerogpu/followups routes follow-up question generation.

Dashboard

The adapter keeps the dashboard idea from the original prototype:

• GET /dashboard/events?limit=50
• GET /dashboard/summary

Runtime events are written to plugin/tracking-events.jsonl. That file is generated on demand and should not be committed.

Configuration

The plugin package includes:

• openclaw.plugin.json for OpenCLAW plugin metadata.
• openclaw-plugin.cjs for provider registration and setup wizard support.
• skills/zerogpu/SKILL.md for OpenCLAW skill metadata.
• scripts/setup-openclaw-provider.sh as a fallback no-clone installer.

Optional environment variables:

• PORT defaults to 8787.
• MODEL_CATALOG_URL defaults to https://api-dashboard.zerogpu.ai/api/models.
• INFERENCE_API_URL defaults to https://api.zerogpu.ai/v1/responses.
• INFERENCE_TIMEOUT_MS defaults to 10000.
• INFERENCE_MAX_RETRIES defaults to 2.
• DEFAULT_SHOW_SAVINGS defaults to true.
• ENABLE_AUTO_TASK_INFERENCE defaults to false. Set true to enable keyword-based task inference for zerogpu/auto.

ZEROGPU_API_KEY and ZEROGPU_PROJECT_ID are only supported as local development fallbacks. Hosted deployments should rely on per-request credentials from OpenCLAW.

Verify

npm run check
bash -n scripts/setup-openclaw-provider.sh
curl http://localhost:8787/health
curl http://localhost:8787/v1/models