agent.js// one fileimport{ tools }async functionrun() {// 25 built-insloop(turn)}$ node agent.js→ ready› describe your agent
DEPS INSTALLED
0

Chat. Build.
Run.

A local-first AI agent harness. Just one file — no setup mess, no cloud account, nothing leaving your laptop. You describe what you want, AI fills it in, and the agent runs on your machine.

1
Single file
0
Things to install
25
Built-in tools
4
AI providers

Agents today mean packages, cloud, and trust you didn't ask for.

Agents today

  • Hundreds of packages just to say hello
  • Hosted-only — your data leaves the machine
  • Runtime you can't read
  • Updates break things quietly

Agent Builder

  • One file that does real work
  • Runs fully on your machine
  • Works offline (with Ollama)
  • Open source, nothing tracked

One loop. Three layers.

01 · User
You describe
02 · Builder
AI assistant fills
03 · Harness
agent.js
Turn loop
Context compression
Subagents + safety gates
Files
read · write
Web
fetch · search
MCP
any server
Sandbox
net off

Everything an agent needs.

25 built-in tools

Files, web, search, knowledge, skills, sandbox — and more. All in the box.

Connect to anything

Plug in tools for GitHub, files, databases, Slack, and dozens more — using the open MCP standard.

Safe sandbox

Risky scripts run in an isolated box — no network, locked to your folder.

Subagents

Big tasks split into smaller ones. Each subagent works on its piece, then reports back.

Skills

Teach the agent new tricks. Drop a folder into ./skills/ and the agent picks it up automatically.

Knowledge that compounds

Pour in articles, papers, links. The agent compiles a wiki that gets richer every run.

Self-extending

Missing a tool? The agent writes one for itself — sandboxed, on the fly.

Pick your AI

Free local AI (Ollama) or cloud (OpenAI, Anthropic, Google). Switch with one flag.

// LOCAL-FIRST

Fully offline. Everything local.

Run Claude Code through Ollama, then run the agents it builds on the same machine. No API keys required for local mode. No cloud account. Nothing leaving your network.

  • Claude Code integration
  • Local or cloud models
  • 64k+ context recommended
  • Plain terminal workflow
ollama launch claude
ollama launch claude --model qwen3.5
→ local model ready
Local
glm-4.7-flash
Cloud
qwen3-coder:480b
64k+
context recommended

More models at ollama.com/search →

Choose your AI assistant.

Any of these can build the agent file. Pick the assistant you like, then point it at the same local harness.

C

Claude Code

Anthropic's agentic coding tool. Use it from the CLI, IDE, desktop app, or browser. Can be launched through Ollama for local model workflows.

Cloud or offline API or Ollama
O

OpenAI Codex CLI

Fast, capable coding sessions for broad agent-building work. Use Codex in the app, CLI, IDE, or cloud, then point it at the same local harness.

Cloud only API or Plus
G

Google Gemini CLI

Useful when you want a Google-account path into agent building. Same prompt loop, same files, same local Agent Builder target.

Cloud only Free tier or API

Clone. Describe.
Run.

01
Open Terminal
Mac · Cmd+Space → Terminal
Windows · Right-click → Git Bash
Linux · Ctrl+Alt+T
02
Clone & set up
git clone builtbyV/agent-builder
cd agent-builder
bash setup.sh
03
Describe your agent
npx claude
› Build me an agent that…