Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/trycua/cua/llms.txt

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

Cua-Bench provides the cb command-line tool for defining, running, and analyzing computer-use benchmark evaluations. Install it from the libs/cua-bench environment, then run cb --help or cb <command> --help to see options for your installed version.

Task and dataset commands

These commands let you inspect, list, scaffold, and build benchmark tasks and dataset artifacts.
CommandPurpose
cb task info <path>Inspect a task’s provider, variants, and lifecycle functions.
cb task list [path]List available task directories under an optional path.
cb task createScaffold a new task interactively.
cb task generateGenerate a task scaffold from a natural-language prompt.
cb dataset listList all known datasets.
cb dataset build <path>Build a dataset artifact from a task directory.

cb task info

Print a task’s metadata: provider, available variants, and which lifecycle hooks are defined.
<path>
string
required
Path to the task directory or task definition file.

cb task list

List available task directories.
[path]
string
Optional root path to search. Defaults to the current working directory.

cb task create

Scaffold a new task interactively. The CLI prompts for a name, provider, and optional variants, then writes the task files.

cb task generate

Generate a task scaffold from a prompt string. Useful for quickly bootstrapping new evaluations.

cb dataset list

List all datasets known to the current Cua-Bench environment.

cb dataset build

Build a dataset artifact from a task directory.
<path>
string
required
Path to the dataset definition or task directory.

Run commands

These commands manage the full benchmark run lifecycle — from launching a single interactive task to running an entire dataset in parallel.
CommandPurpose
cb interact <task-path>Run one task interactively with the VM display visible.
cb run task <task-path>Run a single task, optionally with an agent or oracle.
cb run dataset <dataset>Run all tasks in a dataset, optionally in parallel.
cb run listList all runs and their current status.
cb run info <run-id>Show detailed information for a specific run.
cb run watch <run-id>Stream a run’s progress in the terminal.
cb run logs <run-id>Print logs for a run or session.
cb run stop <run-id>Stop a running benchmark.

cb run task

Run a single benchmark task with an optional agent or oracle. Usage: cb run task <task-path> [options]
<task-path>
string
required
Path to the task definition file or directory.
--agent
string
Agent to use for the run. Built-in values include cua-agent. Pass a custom image name to use a containerized agent.
--model
string
Model identifier to pass to the agent (e.g. gpt-4o, claude-3-5-sonnet).
--oracle
string
Oracle evaluator to use. Overrides agent-based evaluation.
--max-steps
int
Maximum number of agent steps before the task is marked failed.
--image
string
Environment image to use for the task sandbox. Overrides the task’s default image.
--platform
string
Execution platform. See cb platform for available options.
--output-dir
string
Directory to write trajectory files and run artifacts. Defaults to a timestamped directory in the current working directory.

cb run dataset

Run all tasks in a dataset, with optional parallelism. Usage: cb run dataset <dataset> [options]
<dataset>
string
required
Dataset name or path to the dataset definition. Use cb dataset list to find available datasets.
--agent
string
Agent to use for all tasks in the dataset. Built-in: cua-agent. Custom: any agent image name.
--model
string
Model identifier to pass to the agent.
--oracle
string
Oracle evaluator for all tasks.
--max-steps
int
Maximum agent steps per task.
--max-parallel
int
Number of tasks to run concurrently. Higher values require more system resources. Defaults to 1 (sequential).
--image
string
Override the environment image for all tasks.
--platform
string
Execution platform.
--output-dir
string
Directory for trajectory files and run artifacts.

cb interact

Run a single task interactively with the VM display visible. Useful for debugging or manually completing a task. Usage: cb interact <task-path> [options]
<task-path>
string
required
Path to the task to run.
Accepts the same --image, --platform, and --output-dir options as cb run task.

cb run list

List all benchmark runs and their status.

cb run info

Show detailed information about a specific run.
<run-id>
string
required
Run ID from cb run list.

cb run watch

Stream a run’s progress live in the terminal.
<run-id>
string
required
Run ID to follow.

cb run logs

Print logs for a run or a specific session within the run.
<run-id>
string
required
Run ID to fetch logs for.

cb run stop

Stop an active benchmark run.
<run-id>
string
required
Run ID to stop.

Environment and artifact commands

These command groups manage the infrastructure that Cua-Bench tasks run on.

cb image

Manage sandbox environment images used by tasks.
SubcommandPurpose
cb image listList all available environment images.
cb image create <name>Build a new environment image.
cb image delete <name>Delete an environment image.
cb image clone <src> <dst>Clone an image to a new name.
cb image shell <name>Open an interactive shell in an image.

cb image create

Build a new base environment image for use in benchmark tasks.
<name>
string
required
Name for the new image.

cb image list

List all environment images available to Cua-Bench, with their names and statuses.

cb platform

Inspect available execution platforms.
SubcommandPurpose
cb platform listList all configured platforms.
cb platform info <name>Show details for a specific platform.

cb agent

Manage agent images that evaluate tasks.
SubcommandPurpose
cb agent initInitialize a new agent project.
cb agent buildBuild the agent container image.
cb agent pushPush the agent image to a registry.

cb trace

View traces and agent trajectories from completed runs.
SubcommandPurpose
cb trace view <run-id>Open a run’s trace in the terminal viewer.
cb trace open <run-id>Open agent trajectory files from a completed run.

cb status

Inspect currently running Cua-Bench resources (sandboxes, agents, runs).

cb prune

Review and remove unused run artifacts or environment resources.
Use cb prune regularly in CI to reclaim disk space from old trajectory files and stale sandbox images.

Build docs developers (and LLMs) love