cli-and-ui

CLI And UI

Use this sub-skill when the user wants to run smolagents from a terminal or expose an already-designed agent through a local UI/server.

Route First

• For agent construction, CodeAgent vs ToolCallingAgent, model objects, managed agents, and tool definitions, route to the sibling agent/model/tool sub-skills.
• For sandbox executors, remote execution backends, import allow-lists, and security hardening beyond the CLI --imports flag, route to execution-and-safety.
• For this sub-skill, focus on command syntax, UI wrapping, optional extras, browser/UI dependencies, and troubleshooting terminal or serving failures.

Core Entry Points

• smolagent is the general CLI for running a CodeAgent or ToolCallingAgent from a prompt or an interactive setup wizard.
• webagent is a specialized browser-automation CLI that starts a Selenium/Helium Chrome session and runs a vision-capable CodeAgent with screenshot callbacks.
• GradioUI wraps a MultiStepAgent in a streaming Gradio chat interface with optional multimodal file upload.
• launch_gradio_demo(tool) launches a Gradio Interface for a single Tool whose inputs and output_type are defined; use GradioUI for full agent chat UIs.
• Server examples are adaptation patterns: run the blocking agent.run(...) call off the async event loop and disconnect external tool clients during shutdown.

Safe Discovery

Before running model-backed tasks, use safe help checks:

python sub-skills/cli-and-ui/scripts/print_cli_help.py --command smolagent
python sub-skills/cli-and-ui/scripts/print_cli_help.py --command webagent
smolagent --help
webagent --help

If console scripts are unavailable but the package imports, try module-level help:

python -m smolagents.cli --help
python -m smolagents.vision_web_browser --help

Common Workflows

• Run a direct CLI prompt with smolagent "your task" --model-type InferenceClientModel --model-id Qwen/Qwen3-Next-80B-A3B-Thinking --tools web_search.
• Start the guided wizard with bare smolagent; it prompts for action type, tools, model settings, optional API fields, additional imports, and final task.
• Use --action-type code for a code-writing CodeAgent; use --action-type tool_calling for a ToolCallingAgent.
• Use webagent "browser task" --model-type LiteLLMModel --model-id gpt-4o only when browser, model, and optional vision dependencies are installed and a Chrome/WebDriver stack is usable.
• Wrap a configured agent with GradioUI(agent, file_upload_folder="uploads", reset_agent_memory=True).launch(share=False) for a local chat UI.

References

• See CLI reference for command names, flags, defaults, and behavior.
• See UI workflows for GradioUI, streaming, file upload, server adaptation, and monitoring notes.
• See troubleshooting for missing scripts, optional extras, invalid model/tool flags, and browser failures.