AgentLoop
AgentLoop is a minimal, readable AI agent starter built on the Anthropic SDK and Next.js, demonstrating the core streaming tool-use loop in about 150 lines of code. It solves the problem of complex frameworks or bloated templates by providing a simple, understandable foundation for developers who want to build and own their own agent. It is for developers who want to learn how agents work or quickly bootstrap a production-ready agent without a heavy framework.
✨ Key features
- Streaming chat with plain fetch and ReadableStream
- Real tool use loop with two example tools
- Typed and commented Next.js App Router code
- One-command deploy to Vercel
- Free MIT core, no SDK lock-in
🎯 Use cases
- Learn how AI agents work by reading a minimal implementation
- Build a custom agent product with streaming and tool use
- Deploy a simple agent to Vercel quickly
- Extend with custom tools by editing one file
📦 Installation
🧰 Requirements: Node.js, npm, and an Anthropic API key (ANTHROPIC_API_KEY) for running the agent.
git clone https://github.com/mnifzied-create/agentloop.git
cd agentloop
npm install
cp .env.example .env.local # then paste your ANTHROPIC_API_KEY
npm run dev # http://localhost:3000
🚀 Usage
After installation, ask the agent what is (12 * 9) + 7? and watch it call the calculate tool. The agent will stream the response and use tools as needed.
⚠️ Good to know
The free core is a minimal starter with a bounded loop (5 steps) and only two example tools; production features like parallel tool calls, memory, and rate limiting are available in the paid Pro version.
❓ FAQ
How do I build a Claude agent from scratch?
A Claude agent is just a loop: send the conversation to the model, run any tool it calls, append the result, and repeat until it replies. AgentLoop is a free, MIT-licensed starter that does exactly this in about 150 lines.
Why is my AI agent so expensive?
Most agent cost is invisible because you re-send the system prompt and tool schemas on every turn. Verbose tool definitions are billed repeatedly, and a typical support agent carries around 650 tokens of tool schemas per turn.
How do I estimate Claude agent token cost?
Count what you re-send each turn (system prompt, tool schemas, prior messages, tool outputs) and multiply by turns and model price. The free Agent Token Profiler does this in your browser.
How do I reduce my AI agent's token cost?
Trim verbose tool schemas, summarize tool outputs, cap conversation history, and route easy turns to a cheaper model like Claude Haiku. Measure first with the Token Profiler.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.