Pyvello

A Python-First Reactive AI UI Framework with React 18, shadcn/ui, LiteLLM Agent Swarms, and Automatic Graph Visualization

Write pure Python. Get an enterprise-grade React 18 UI with Tailwind CSS and Radix UI primitives. Build autonomous multi-agent swarms with LiteLLM, attach tools (RAG, SQL, Sheets, Gmail), connect self-hosted inference clusters (vLLM, Ollama), and observe execution in real-time over persistent WebSockets.

Features

• Python-First Simplicity: Build full web applications without writing React, JSX, or CSS.
• shadcn/ui & 18 Calibrated Palettes: Coherent design tokens, dark and light modes, accessible Radix UI primitives, clean typography.
• LiteLLM Multi-Agent Swarms: Implement all 6 Anthropic patterns (Prompt Chaining, Routing, Parallelization, Orchestrator-Workers, Evaluator-Optimizer, and Autonomous Multi-Tool) across 100+ models.
• Custom Inference Base URLs: Route requests to self-hosted vLLM clusters (http://localhost:8000/v1), local Ollama (http://localhost:11434/v1), LM Studio, or OpenRouter via api_base.
• Builtin Enterprise Tools: Out-of-the-box, fully runnable tools for RAG vector search, relational SQL databases, Google Sheets sync, and Gmail dispatch.
• Universal API Client (pyvello.api): Axios-style REST client, SSE streaming, and WebSockets with automatic component loading state orchestration (with api.loading(btn, card): ...).
• Real-Time Agent Graph Visualization: Live interactive canvas powered by @xyflow/react and Dagre auto-layout. Watch nodes execute with live duration tracking, token telemetry, and slide-over inspectors.
• Production DataTable: Sorting, global search, pagination, and 3-dots action menus with conditional visibility rules.
• Interactive SVG Charts: Area charts, bar charts, line charts, and donut charts with responsive tooltips.
• Zero-Node Runtime with Instant Hot Reloading: Production React bundle is packaged directly inside the Python wheel. Hot reload is enabled by default for rapid local development.
• Jupyter Notebook & Google Colab Support: Embed responsive full-stack apps inline inside .ipynb cells via non-blocking background threads.
• LangGraph-Style Conditional Tool Nodes: Dedicated tool nodes connected with conditional call and result return edges, rendered on a full-screen canvas.

Quickstart

1. Installation

pip install pyvello

All dependencies (FastAPI, Uvicorn, LiteLLM, HTTPX, WebSockets, Pydantic, Watchfiles) install automatically.

2. Autonomous Multi-Tool Agent (RAG, SQL, Sheets, Gmail)

import pyvello as ui
from pyvello.tools import (
    query_knowledge_base,
    query_database,
    read_google_sheet,
    send_gmail_message,
)

# Initialize Agent with Tools, Model Switching, and Custom Base URL
agent = ui.agent(
    name="EnterpriseOpsAgent",
    model="gpt-4o",
    models=[
        {"id": "gpt-4o", "name": "GPT-4o (OpenAI)", "badge": "Cloud"},
        {"id": "claude-3-5-sonnet", "name": "Claude 3.5 Sonnet", "badge": "Coding"},
        {"id": "meta-llama/Llama-3.3-70B-Instruct", "name": "Llama 3.3 (Local vLLM)", "badge": "Local GPU"},
    ],
    system_prompt="You are an enterprise operations assistant. Use RAG for policies, SQL for DB queries, Sheets for CRM deals, and Gmail to dispatch summaries.",
    tools=[query_knowledge_base, query_database, read_google_sheet, send_gmail_message],
    temperature=0.2,
    api_base="http://localhost:8000/v1",  # Self-hosted vLLM / Ollama endpoint
)

# Launch interactive UI with one call
agent.to_app().launch()

3. Multi-Agent Swarm Orchestrator

import pyvello as ui

# 1. Specialized Worker Subagents
market_worker = ui.agent(name="MarketResearcher", system_prompt="Analyze target addressable market and unit economics.")
infra_worker = ui.agent(name="InfraArchitect", system_prompt="Specify cloud infrastructure and multi-region Kubernetes nodes.")
risk_worker = ui.agent(name="RiskAssessor", system_prompt="Evaluate compliance, HIPAA guidelines, and operational risk.")

# 2. Dynamic Swarm Orchestrator with Supervisor System Prompt & Model Selection
orchestrator = ui.orchestrator(
    agents=[market_worker, infra_worker, risk_worker],
    supervisor_system_prompt="You are an enterprise software architect supervising junior engineers. Decompose goals and synthesize verified deliverables.",
    supervisor_model="gpt-4o",
    models=[
        {"id": "gpt-4o", "name": "GPT-4o (Supervisor)", "badge": "Cloud"},
        {"id": "claude-3-5-sonnet", "name": "Claude 3.5 Sonnet", "badge": "Coding"},
        {"id": "meta-llama/Llama-3.3-70B-Instruct", "name": "Llama 3.3 (vLLM :8000)", "badge": "Local GPU"},
    ],
    temperature=0.2,
    stream=True,
    structured_output={"type": "json_object"},
    api_base="http://localhost:8000/v1",  # Cascades to workers automatically!
)

# 3. Mount in Reactive Chat UI
orchestrator.to_app().launch()

4. Universal API Client & Automatic Loading States

from pyvello import api, ui

client = api.ApiClient(base_url="https://api.mycompany.org/v1")

with ui.App("Telemetry Dashboard"):
    m1 = ui.metric("Active Sessions", "14,280")
    btn = ui.button("Sync Telemetry", icon="RefreshCw", loader_style="spinner")

    # Automatically sets button loading=True, renders spinner, and restores on complete
    @btn.on("click")
    def on_sync():
        with api.loading(btn, m1):
            res = client.get("/telemetry/metrics")
            ui.update(m1, value=res.json()["active_sessions"])
            ui.toast("Metrics synchronized!", "success")

5. Custom Tools with Automatic JSON Schema Extraction

import pyvello as ui

@ui.tool(description="Calculate portfolio Sharpe ratio given returns and risk-free rate")
def calculate_sharpe(returns: list[float], risk_free_rate: float = 0.04) -> float:
    mean_ret = sum(returns) / len(returns)
    variance = sum((r - mean_ret) ** 2 for r in returns) / len(returns)
    std_dev = variance ** 0.5
    return (mean_ret - risk_free_rate) / std_dev if std_dev > 0 else 0.0

agent = ui.agent(
    name="PortfolioAnalyst",
    tools=[calculate_sharpe],
    model="gpt-4o"
)

All 6 Agent Architecture Patterns Supported

18 Calibrated Theme Palettes

Red (palette="red") is configured as the framework default theme palette. You can pass any of the 18 calibrated theme palettes to ui.App, ui.ChatInterface, or agent.to_app():

red (default), rose, violet, indigo, blue, sky, cyan, teal, emerald, green, yellow, amber, orange, zinc, slate, stone, gray, neutral.

Toggle dark and light modes with persistent client preference syncing.

CLI & Notebook Reference

# Start local dev server (Hot reload enabled by default)
pyvello run app.py

# Specify port or disable reload if needed
pyvello run app.py --port 8000 --no-reload

# Export application to a standalone React 18 + Vite 5 project
pyvello export app.py --output dist/

Jupyter Notebook (.ipynb) Inline Embedding

import pyvello as ui

agent = ui.agent(name="Analyst", system_prompt="You are a data analyst.")

# Embeds directly into the notebook output cell without blocking kernel:
agent.to_app().launch(inline=True, height=750)

License

MIT License.