Documentation
Everything you need to build, customize, and ship browser-executable AI agents.
Getting Started
Create your first agent in five minutes, no setup needed.
Building Agents
Python tools, tool schemas, async handlers, and code structure.
Templates
Create reusable agent blueprints with custom UI and prompts.
AI Providers
OpenAI, Anthropic, or fully local wllama — compare and configure.
API Reference
Agent properties, template variables, and the HTML generation flow.
FAQ
Costs, privacy, local models, debugging — common questions answered.
Quick start
Sign up for free
Create an account by signing up — no credit card required. You can start building immediately with local models that need no API key.
Pick a starting point
Choose a template from the Create page — or click "Use this template" from the marketplace. Each one ships with a Python scaffold, a UI, and a sample system prompt — ready to run.
Try it, then download
Click "Try it now" on your agent's page to chat with it right in the browser. Happy with it? Download a standalone HTML file — the model loads, your Python executes, and the agent runs entirely on the visitor's machine.
Customize (optional)
Write your own async def tool functions and export their schemas via get_tool_schemas() — or leave the template defaults as-is. AgentOp runs everything in-browser via Pyodide, no server required.
What is AgentOp?
AgentOp lets you build AI agents that run as a single standalone HTML file — no backend, no server, no infra. Your Python tools execute in-browser via Pyodide, and the LLM runs either via the OpenAI/Anthropic APIs or locally via wllama (llama.cpp compiled to WebAssembly). Prompts and data stay on the user's device.
Key concepts
- Agents — A named project with Python code, a template, a provider, and optional packages. Downloads as a single HTML file.
- Templates — Reusable blueprints that define the UI (HTML/CSS/JS) and prompt configuration. Agents are created from templates.
- Tool functions —
async defPython functions your agent can call (search, calculate, query a database, etc.). Defined once, dispatched automatically by the LLM. - Providers — Where the LLM runs: OpenAI GPT-4o, Anthropic Claude, or a local wllama model that downloads to the user's browser.