I wanted a way to run an AI agent loop without installing anything, managing dependencies, or paying for a SaaS subscription. So I built one that runs entirely in your browser.
AI Agent Builder — Free, BYOK, No Backend
What it does
Paste your OpenAI or Gemini API key, type a goal, and watch an autonomous agent work through it step by step. The agent plans, executes, and verifies — all client-side. Your key goes directly to the API provider, never to my servers. There is no backend.
Four strategies based on real research
The tool implements four proven agentic loop patterns from Anthropic, OpenAI, and Google DeepMind research:
Plan & Execute — breaks the goal into steps, executes each one, then verifies. Best for structured tasks like research or content creation.
ReAct (Reason + Act) — the agent thinks, acts, observes, and adapts on every iteration. From the original Princeton/Google Brain paper. Best for open-ended exploration where you don't know the steps upfront.
Reflection Loop — generates a draft, critiques it, then improves. Repeats up to 3 times. Based on Reflexion (Northeastern/MIT) which hit 91% on HumanEval. Best for writing and code.
Fast Mode (ReWOO) — plans everything in one call, executes in one call. Just 2 API calls total. Best for simple tasks when you want to save money.
The build
One HTML file. 54KB. Vanilla JavaScript. No React, no build step, no npm. Hosted on Cloudflare Pages free tier.
The hardest part was getting the ReAct loop to parse reliably — the model needs to output structured "Thought → Action → Observation" blocks, and different models format them differently. I ended up parsing for "Final Answer:" as the termination signal, which works across GPT-4o and Gemini.
Cost tracking
Every API call shows token usage and estimated cost in real-time. A typical 5-step agent run costs $0.01-0.05 with GPT-4.1 Nano. You can watch the cost accumulate and kill the run if it gets expensive.
Try it
billion-agentbuilder.pages.dev
Give it something like "Research the pros and cons of Rust vs Go for backend development" and watch it work. The ReAct strategy is the most fun to watch — you can see the agent reasoning in real time.
Curious which strategy works best for your use case. I've been defaulting to Plan & Execute but the Reflection loop produces noticeably better writing.
Top comments (0)