Autonomous Building

Autonomous Application Building

This skill teaches agents how to use PyWry MCP's agentic tools to autonomously design, build, and export complete widget applications from a plain-English description.

When to Use These Tools

Use the agentic tools when you need to:

• Build a complete PyWry app from a description without manual step-by-step tool calls
• Deliver a full Python project (main.py, requirements.txt, README.md, widgets/)
• Interactively gather requirements from the user before generating code
• Show real-time progress to the user during long build operations

Available Agentic Tools

plan_widget — AI-Planned Widget Spec

Sends the description to the LLM via sampling and returns a validated WidgetPlan JSON object. Use this to inspect the plan before committing to building.

# Returns a WidgetPlan JSON (no widget is created yet)
result = await client.call_tool("plan_widget", {
    "description": "A crypto price dashboard with symbol selector and refresh button"
})
plan = json.loads(result[0].text)
# plan contains: title, html_content, toolbars, callbacks, width, height, ...

build_app — End-to-End Autonomous Builder

The primary tool for autonomous app building. One call:

1. Samples a WidgetPlan from the description
2. Registers the widget in the session
3. Returns widget_id and complete runnable Python code

result = await client.call_tool("build_app", {
    "description": "Task tracker with add/remove buttons and completion percentage",
    "open_window": False  # set True to open a native window immediately
})
data = json.loads(result[0].text)
widget_id  = data["widget_id"]
python_code = data["python_code"]  # paste into a .py and run directly

The returned python_code is a fully self-contained Python script requiring only pywry.

export_project — Complete Python Project Package

Takes one or more widget_ids and generates a full project tree:

my_app/
    main.py            ← entry-point
    requirements.txt   ← dependencies
    README.md          ← quickstart docs
    widgets/
        <widget_id>.py ← one file per widget

result = await client.call_tool("export_project", {
    "widget_ids": ["abc123", "def456"],
    "project_name": "my_dashboard",
    "output_dir": "",          # leave empty to get file contents as JSON
    # "output_dir": "/tmp"     # set to write files to disk
})
data = json.loads(result[0].text)
files = data["files"]  # {relative_path: file_content}

scaffold_app — Interactive Multi-Turn Builder

Uses ctx.elicit() to ask the user questions before generating the plan:

• App title and description
• Display mode (native / inline)
• Optional libraries (Plotly, AG-Grid)
• Toolbar position

result = await client.call_tool("scaffold_app", {})
# MCP client will prompt the user for each field
data = json.loads(result[0].text)
plan = data["widget_plan"]

Recommended Workflow

Quick build (autonomous)

# 1. Build the app
build = await client.call_tool("build_app", {
    "description": "Your plain-English description here"
})
data = json.loads(build[0].text)

# 2. Save the code
Path("my_widget.py").write_text(data["python_code"])

# 3. Or package as a full project
project = await client.call_tool("export_project", {
    "widget_ids": [data["widget_id"]],
    "project_name": "my_app",
    "output_dir": "./output"  # writes files to disk
})

Inspect-then-build

# 1. Plan first
plan_result = await client.call_tool("plan_widget", {
    "description": "..."
})
plan = json.loads(plan_result[0].text)

# 2. Review and tweak the JSON plan manually
# 3. Then build using the reviewed description

Interactive (with user input)

# Let the user guide the design
scaffold = await client.call_tool("scaffold_app", {})
data = json.loads(scaffold[0].text)

# Build from the collected spec
build = await client.call_tool("build_app", {
    "description": data["collected"]["description"]
})

Progress Reporting

All agentic tools emit report_progress events. Clients that display a progress bar will show real-time status: Planning… → Generating code… → Writing files… → Done

Combining With Other Tools

After build_app you can continue refining using the standard tools:

# Read the skill resource for styling tips
content = await client.read_resource("skill://styling/SKILL.md")

# Update a component dynamically
await client.call_tool("set-content", {
    "widget_id": widget_id,
    "component_id": "main-content",
    "content": "<p>Updated!</p>"
})

# Check events emitted by toolbar buttons
await client.call_tool("get-events", {"widget_id": widget_id})

Tips for Good Results

• Specificity beats brevity: "A real-time stock ticker with 5 company buttons and a price display area" produces better plans than "a finance app".
• Mention coloring / layout: "left sidebar with dark theme", "top toolbar with primary variant buttons".
• Describe interactivity: "clicking a country updates the chart below" triggers better callback planning.
• Large widgets: Request include_plotly=True or include_aggrid=True in your description if you need charts or tables.