CLI Reference¶

OpenAdapt provides a unified command-line interface for all functionality.

Global Commands¶

Version¶

Show installed package versions:

openadapt version

Output:

openadapt: 1.0.0
openadapt-capture: 1.0.0
openadapt-ml: 1.0.0
openadapt-evals: 1.0.0
openadapt-viewer: 1.0.0

Doctor¶

Check system requirements and configuration:

openadapt doctor

This verifies:

• Python version
• Required packages
• System permissions (macOS)
• GPU availability
• Environment variables

Capture Commands¶

Commands for collecting human demonstrations.

capture start¶

Start a new demonstration collection session.

openadapt capture start --name <name> [options]

Arguments:

Examples:

# Basic demonstration collection
openadapt capture start --name login-task

# Demonstration collection without screenshots
openadapt capture start --name audio-task --no-screenshots

# Demonstration collection with slower screenshot interval
openadapt capture start --name slow-task --interval 1.0

capture stop¶

Stop the current demonstration collection.

openadapt capture stop

Alternatively, press Ctrl+C in the capture terminal.

capture list¶

List all captured demonstrations.

openadapt capture list

Output:

NAME           EVENTS   DURATION   DATE
login-task     45       2m 30s     2026-01-16
email-reply    23       1m 15s     2026-01-15
form-fill      89       5m 42s     2026-01-14

capture view¶

Open the trajectory viewer for a demonstration.

openadapt capture view <name> [options]

Arguments:

capture delete¶

Delete a demonstration.

openadapt capture delete <name>

Train Commands¶

Commands for policy learning from demonstrations.

train start¶

Start policy learning from a demonstration.

openadapt train start --capture <name> --model <model> [options]

Arguments:

Available Models:

• qwen3vl-2b - Qwen3-VL 2B parameters
• qwen3vl-7b - Qwen3-VL 7B parameters
• llava-1.6-7b - LLaVA 1.6 7B parameters

Examples:

# Basic policy learning
openadapt train start --capture login-task --model qwen3vl-2b

# Policy learning with custom parameters
openadapt train start \
    --capture login-task \
    --model qwen3vl-7b \
    --epochs 20 \
    --batch-size 2 \
    --learning-rate 5e-5

train status¶

Check policy learning progress.

openadapt train status

Output:

Training: login-task
Model: qwen3vl-2b
Progress: Epoch 5/10 (50%)
Loss: 0.234
ETA: 15 minutes

train stop¶

Stop the current policy learning.

openadapt train stop

train models¶

List available model architectures.

openadapt train models

Eval Commands¶

Commands for evaluating agents.

eval run¶

Run an evaluation.

openadapt eval run [options]

Arguments:

*One of --checkpoint or --agent is required.

Available Benchmarks:

• waa - Windows Agent Arena
• osworld - OSWorld
• webarena - WebArena
• mock - Mock benchmark for testing

Examples:

# Evaluate a trained model
openadapt eval run --checkpoint training_output/model.pt --benchmark waa

# Evaluate Claude API agent
openadapt eval run --agent api-claude --benchmark waa

# Run subset of tasks
openadapt eval run --agent api-claude --benchmark waa --tasks 10

eval mock¶

Run a mock evaluation to test setup.

openadapt eval mock --tasks <n>

Arguments:

eval benchmarks¶

List available benchmarks.

openadapt eval benchmarks

Serve Commands¶

Start the dashboard server.

openadapt serve [options]

Arguments:

Access the dashboard at http://localhost:8080.

Command Structure¶

graph LR
    OA[openadapt]

    OA --> CAP[capture]
    OA --> TRN[train]
    OA --> EVL[eval]
    OA --> SRV[serve]
    OA --> VER[version]
    OA --> DOC[doctor]

    CAP --> CS[start]
    CAP --> CT[stop]
    CAP --> CL[list]
    CAP --> CV[view]
    CAP --> CD[delete]

    TRN --> TS[start]
    TRN --> TST[status]
    TRN --> TSP[stop]
    TRN --> TM[models]

    EVL --> ER[run]
    EVL --> EM[mock]
    EVL --> EB[benchmarks]

Environment Variables¶

Exit Codes¶