TypeClawQuickstart
Install TypeClaw, start it, and get your first reply — in about five minutes
By the end of this page you'll have a TypeClaw agent running on your machine and answering you in a terminal. Five steps, nothing to configure ahead of time. The deeper "why" for each step is tucked into the expandable boxes — skip them on your first pass.
Before you start
You need two things installed:
- Bun (version 1.1 or newer) — the runtime TypeClaw is built on.
- OrbStack (recommended) or Docker — your agent runs inside a container. OrbStack is lighter and faster to start on macOS; Docker Desktop works just as well.
You'll also need access to one AI provider — an API key from OpenAI, Anthropic, Fireworks, Z.AI, xAI, MiniMax, or DeepSeek, or an existing ChatGPT, Claude, or Grok subscription you can log in with (via the openai-codex, anthropic, or xai providers). The wizard in step 2 walks you through it.
Five steps to your first reply
Create your agent and run the setup wizard
Pick a name for your agent — the folder name becomes the agent's name — then run init inside it:
mkdir my-agent && cd my-agent
typeclaw initinit is an interactive wizard. It asks which AI provider you want, which model, and how to sign in (an API key, or a browser login if you pick a subscription-backed provider like OpenAI Codex or Anthropic) — and it offers to wire up your first channel (Slack, Discord, GitHub, …) right then too. Answer the prompts and it sets everything up for you. By the time init finishes, you have a working provider and, if you said yes, a channel already connected.
Start the agent
typeclaw startThe first run builds your agent's container, which takes a few minutes — you'll see Docker build output scroll past. Later starts take seconds. When it's done, your agent is running in the background.
Talk to it
typeclaw tuiThis opens a terminal chat window connected to your agent. Type a message, press enter, and watch it respond. That's the whole loop.
Want a single quick answer without opening the chat UI?
typeclaw tui "hello, who are you?"Stop it when you're done
typeclaw stopThat cleanly shuts the container down. To start again later, just run typeclaw start from the same folder.
What just happened
Your agent runs inside its own Docker container. The typeclaw commands you ran — start, stop, tui — don't contain the agent; they're just a remote control that talks to the container. The agent can't reach the rest of your computer: its world is the agent folder you just created plus the container. The host CLI itself never runs the agent loop or loads its plugins — so the agent's state stays in that one folder and its container, not scattered across your machine. That separation is the core idea behind TypeClaw, and it's what makes it safe to give the agent real work. The full picture is in How it's built.
Next: Teach the agent — it gets sharper the more you use it, on its own.