Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.hud.ai/llms.txt

Use this file to discover all available pages before exploring further.

Grounding tools convert natural language element descriptions to pixel coordinates. Agent says “click the red submit button”—grounder finds it and returns coordinates.

​
How It Works

Agent: "click the red submit button"
         ↓
    [Screenshot]
         ↓
    [Vision Model: (450, 320)]
         ↓
    Computer: click(x=450, y=320)

​
GroundedComputerTool

Wraps a computer tool to accept element descriptions instead of coordinates. 
from hud.tools.grounding import GroundedComputerTool, Grounder, GrounderConfig

config = GrounderConfig(
    api_base="https://api.openai.com/v1",
    model="gpt-4o",
    api_key="your-api-key",
)
grounder = Grounder(config=config)

grounded = GroundedComputerTool(
    grounder=grounder,
    ctx=env,                        # Environment context
    computer_tool_name="computer",  # Name of computer tool to use
)
Actions: click, double_click, move, scroll, drag, type, keypress, screenshot, wait 
# Click using description
await grounded(
    action="click",
    element_description="the blue login button at the top",
    screenshot_b64=current_screenshot,
)

# Scroll at element
await grounded(
    action="scroll",
    element_description="the main content area",
    scroll_x=0,
    scroll_y=-100,
    screenshot_b64=current_screenshot,
)

# Drag between elements
await grounded(
    action="drag",
    start_element_description="the file icon",
    end_element_description="the trash folder",
    screenshot_b64=current_screenshot,
)

# No grounding needed for these
await grounded(action="type", text="Hello!")
await grounded(action="keypress", keys=["ctrl", "s"])
Screenshot is required for actions that need grounding.

​
Grounder

The engine that locates elements using vision models. 
from hud.tools.grounding import Grounder, GrounderConfig

# Basic config
config = GrounderConfig(
    api_base="https://api.openai.com/v1",
    model="gpt-4o",
)
grounder = Grounder(config=config)

# With custom settings
config = GrounderConfig(
    api_base="https://openrouter.ai/api/v1",
    model="qwen/qwen-2.5-vl-7b-instruct",
    api_key="your-openrouter-key",
    output_format="pixels",
)
grounder = Grounder(config=config)

coords = await grounder.predict_click(
    image_b64=screenshot_base64,
    instruction="the submit button",
)
# Returns: (x, y) or None if not found
Supported models: Any vision-capable model via OpenAI-compatible API—GPT-4o, Qwen VL, LLaVA, etc.

​
With HUD Agents

GroundedComputerTool is typically used as a wrapper around environment computer tools. Register the underlying computer tool, then use grounded calls: 
from hud import Environment
from hud.tools import AnthropicComputerTool
from hud.tools.grounding import GroundedComputerTool, Grounder, GrounderConfig

# Setup environment with computer tool
env = Environment("grounded-env")
env.add_tool(AnthropicComputerTool())

# Create grounder
config = GrounderConfig(
    api_base="https://api.openai.com/v1",
    model="gpt-4o",
    api_key="your-api-key",
)
grounder = Grounder(config=config)

async with env:
    # Wrap environment for grounded calls
    grounded = GroundedComputerTool(grounder=grounder, ctx=env)
    
    # Take screenshot via environment
    result = await env.call_tool("computer", action="screenshot")
    
    # Use grounded tool for element-based actions
    await grounded(
        action="click",
        element_description="the login button",
        screenshot_b64=result.content[0].data,  # base64 from screenshot
    )
For full agent loops, use HUD’s built-in agents which handle the loop automatically: 
from hud.agents import create_agent
import hud

task = env("my_task")
agent = create_agent("gpt-4o")

async with hud.eval(task) as ctx:
    await agent.run(ctx)

​
When to Use

Good for:
  • Dynamic interfaces where elements move
  • Natural language task descriptions
  • Complex layouts with many similar elements
Avoid when:
  • Static, known positions
  • High-frequency actions (grounding adds latency)
  • Precision required (coordinates are more exact)

​
Trade-offs

AspectGroundedDirect Coordinates
FlexibilityHighLow
PrecisionMediumHigh
SpeedSlowerFaster
Error handlingDescriptiveSilent failures

​
Tips

Write specific descriptions. “The blue submit button at the bottom of the form” beats “the button”. Always use recent screenshots. Stale images lead to wrong coordinates if UI changed. Handle None returns. Grounder returns None if it can’t find the element—provide fallback behavior. → Computer Tools — Underlying computer control