GoClick on RunPod Serverless — low-cost setup

Endpoint settings

Use these settings for about 500 requests/day:

• Endpoint type: Queue
• GPU category: 16 GB
• Active/minimum workers: 0
• Maximum workers: 1
• GPUs per worker: 1
• Idle timeout: 5 seconds
• Execution timeout: 30 seconds
• FlashBoot: Enabled
• Model: HongxinLi/GoClick-Large
• Environment variable: MODEL_NAME=HongxinLi/GoClick-Large
• Do not add a paid network volume only for this model
• Use RunPod Model Caching; do not bake model weights into the image

For the lowest bill, replace both values with HongxinLi/GoClick-Base. It loads and runs faster, but accuracy is lower.

Deploy

1. Push this folder to GitHub.
2. Open RunPod → Serverless → New Endpoint.
3. Deploy from the GitHub repository using this Dockerfile.
4. Select Queue endpoint and the settings above.
5. In the endpoint Model field, enter HongxinLi/GoClick-Large.
6. Deploy.

The worker finds the cached model under /runpod-volume/huggingface-cache/hub and loads it in offline mode.

Call it

Compress screenshots to JPEG quality 65–80 to reduce request size and transfer time.

export RUNPOD_ENDPOINT_ID="your-endpoint-id"
export RUNPOD_API_KEY="your-api-key"
node call-endpoint.mjs screenshot.jpg "the Send button at the bottom-right"

RobotJS scaling

const screen = robot.getScreenSize();
const clickX = Math.round(output.x * screen.width / output.image_width);
const clickY = Math.round(output.y * screen.height / output.image_height);
robot.moveMouseSmooth(clickX, clickY);
robot.mouseClick();

Cost notes

• 500 requests in bursts across a few agent sessions is inexpensive because one warm worker handles many requests.
• 500 isolated requests spread through the day can be much more expensive because each may cold-start.
• Keep max workers at 1 initially.
• Keep idle timeout at 5 seconds initially.
• Never set active workers to 1 for this traffic level.