Get started
// docs

Getting Started

Create your first AI agent in minutes — no backend, no API key needed to start.

Welcome to AgentOp — Build AI Agents That Run in the Browser

AgentOp is a platform for creating browser-executable AI agents using Python. An AgentOp agent is a standalone HTML file that runs a complete LLM pipeline entirely inside the browser — no backend server, no API proxy, no infrastructure to manage. You write Python tool functions, choose an AI provider (OpenAI GPT-4o, Anthropic Claude, or a fully local wllama model), and AgentOp bundles everything into a single downloadable HTML file. Open it in any modern browser and your agent runs instantly.

Under the hood, AgentOp uses Pyodide (Python compiled to WebAssembly) to execute your Python code client-side, and either the OpenAI/Anthropic APIs or wllama (llama.cpp compiled to WebAssembly) to power the language model. The result is a self-contained agent you can host on any static CDN, email as an attachment, or run completely offline after the first load.

This guide walks you through creating your first agent, from account setup to deployment.

Prerequisites

Before you begin, you'll need:

  • A modern web browser (Chrome, Firefox, Safari, or Edge)
  • An AgentOp account (free to create)
  • Optional: API keys for OpenAI or Anthropic (if not using local models)

💡 No AI API keys needed to get started!

You can use local wllama models that run entirely in your browser without any API costs.

Step 1: Create an Account

Start by signing up for a free AgentOp account:

  1. Visit the signup page
  2. Enter your username, email, and password
  3. Click "Sign Up" to create your account
  4. You'll be automatically logged in

Step 2: Create Your Agent

Click Create in the navigation menu and pick a starting point — a template with a working UI, example Python code, and a sample system prompt already wired up. Give your agent a name and a short description, then click "Create Agent."

💡 What's a template?

A template is a ready-made agent design. Creating an agent gives you your own copy to customize and download. You can browse templates first in the marketplace and click "Use this template" to jump straight into creating — or just pick one from the grid on the Create page.

By default your new agent uses a local model (wllama) — it runs entirely in your browser, with no API key and no cost. You can switch to OpenAI or Anthropic later if you want cloud models.

Step 3: Try It in Your Browser

You'll land on your agent's detail page. Click "Try it now" to run it right there in a sandboxed preview — no download needed yet:

  • Local (wllama): The model downloads from HuggingFace on first run (downloads range from ~0.5GB to ~7.5GB; the recommended Qwen 3 4B is ~2.6GB) and is cached in your browser afterwards
  • OpenAI/Anthropic: You'll be asked for your API key, which stays in your browser and is never sent to AgentOp's servers
  • Type a message and press Enter to chat
  • If your agent has tool functions, try triggering them with relevant requests

Step 4: Download or Share

Once you're happy with how it behaves:

  1. Click "Download" to get a standalone HTML file you can save, email, or host anywhere
  2. If using OpenAI or Anthropic, you'll be prompted to encrypt your API key before it's embedded in the file
  3. Open the downloaded file in any browser to run it completely offline

🔒 Security Note: API Keys

API keys are encrypted client-side before being embedded in the HTML file. For maximum security, use a separate API key with usage limits for each agent.

Want others to find it? Edit your agent and set "Visibility" to Public — it'll appear in the marketplace for others to discover, download, and fork.

Step 5: Make It Yours (Optional)

Everything below is optional — your agent works fine with template defaults. Open "Edit" on your agent's detail page to customize:

  • Instructions: Give your agent its own personality — expand "Personality & instructions" on the create page, or the "Instructions" field when editing. Leave blank to inherit the template's default instructions.
  • AI Provider: OpenAI (GPT-4o-mini/GPT-4o), Anthropic (Claude), or stick with local wllama
  • Python code: Tool functions (async def) your agent can call, and the get_tool_schemas() function that describes them
  • Packages: Pyodide built-ins (numpy, pandas, matplotlib) or any PyPI package via micropip
  • Tags and visibility: Categorize your agent and choose whether it's public or private

💡 Prompts, mostly, are configured separately

System prompts and user prompt templates come from the template by default — not your Python code — with the Instructions field above letting you override just the system prompt per agent. AgentOp handles all LangChain orchestration, memory, and streaming automatically.

Deploying the HTML File Elsewhere

  • Host on a website: Upload to your web server
  • GitHub Pages: Create a repository and enable Pages
  • Email distribution: Send the HTML file directly to users
  • Embed in applications: Include via iframe or as a component

Next Steps

Now that you've created your first agent, explore more advanced features:

Building Agents →

Deep dive into Python tools, tool schemas, and async handlers

Template System →

For creators: build your own reusable agent blueprints

AI Providers →

OpenAI, Anthropic, and local wllama setup

API Reference →

Technical reference for advanced users

Need Help?

If you run into issues:

  • Check the FAQ for common problems
  • Browse example agents in the marketplace
  • Fork working agents to learn from their code