OpenClaw Full Tutorial for Beginners: How to Setup Your First AI Agent (ClawdBot)

What the video covers (high level)

• End-to-end beginner walkthrough: install OpenClaw from scratch, deploy safely, and get practical first‑day usage tips so you don’t waste tokens or expose secrets.
• Demo uses a Raspberry Pi, but advice also applies to a Mac Mini, old laptop, or VPS.

Key concepts, features, and workflow

Recommended deployment approach

• Use a dedicated device (not your main machine with personal files/passwords).
• Prefer starting from a blank slate (manual install) rather than vendor one‑click images — preconfigured images can limit later customization.
• MVP-first: install a minimal configuration, use the agent for a few days to learn needs, then wipe and reconfigure deliberately.

“Start small, learn how the agent behaves, then iterate — don’t deploy many agents immediately.”

Installation flow (terminal commands)

• Run the provided install script: it checks prerequisites, downloads dependencies, installs OpenClaw, and starts onboarding.
• During onboarding choose:
  • Manual onboarding
  • Local gateway
  • Loopback binding
  • Generate token
  • Optionally disable Tailscale
• Enable gateway/service, node, and bash as prompted.

API model provider recommendation

• Use OpenRouter as a unified API provider: one API key can access many models (OpenAI, Google, Anthropic, and open‑source models).
• Start with a cheaper open‑source daily driver model (demo subtitles referenced a model like “Miniax”) for low‑cost experimentation. Subtitle price comparisons may be approximate — verify current pricing before large usage.

Messaging / channel setup

• Configure a Telegram bot via BotFather: create the bot and copy the bot token into OpenClaw.
• Use an allowlist (specific Telegram user IDs) to restrict who can DM the agent.
• After install, a dashboard URL with a token is shown for direct testing.

Agent bootstrapping and Markdown (MD) file system

• The first conversation bootstraps the agent’s identity and purpose.
• OpenClaw stores configuration and context in Markdown (.md) files, for example:
  • bootstrap.md — first‑run onboarding prompt
  • agents.md — instruction set for the agent
  • identity.md / persona — agent identity/personality
  • user.md — who you are: name, timezone, communication preferences, projects
• Agents can read and edit these files. You can evolve agent behavior by updating files or instructing the agent to update its own MD files.

Security guidance & automated fixes

• Run the agent’s built‑in security audit and health checks; the agent can scan the host and recommend/harden configuration.
• Enable log redaction to avoid leaking API keys or sensitive metadata in conversation logs. The agent can enable and apply these protections.
• Running the security audit and enabling redaction significantly improves deployment safety in the demo.

Web search and external data

• Configure web search using Brave Search or Perplexity.
  • Brave Search: free credits, roughly ~1 request/sec and ~2,000 requests/month on the demo plan (check current limits).
  • Generate an API key and paste it into onboarding or use the CLI helper to enable web fetch.
• Test the web search from the agent to confirm functionality.

Automations and daily tasks

• Use the agent to schedule regular automations (example: daily AI news at 8 a.m.).
• Start small, run automations for several days, and iteratively refine scope and output.
• Agents can optimize their own tasks over time when instructed.

Best practices for the first days

• Don’t spin up many agents immediately. Focus on one primary agent.
• Voice‑note / brain‑dump goals and let the agent populate MD files (e.g., research.md, travel.md) with tasks and notes.
• After building out MD files, spawn specialized agents (research, writing, etc.) seeded from those files for faster setup and fewer wasted tokens.

Tools / services mentioned

• OpenClaw (agent framework)
• OpenRouter (unified API access to models)
• OpenAI, Google, Anthropic models
• Open‑source models (cheap alternatives; demo references “Miniax”)
• Claude, Opus (comparisons)
• Telegram Bot (BotFather)
• Brave Search API (or Perplexity as an alternative)
• Raspberry Pi (demo device), Mac Mini, VPS, old laptops (deployment options)

Practical step list (condensed)

1. Use a dedicated device and open a terminal on that device.
2. Run the installer script from the video/source to perform prerequisite checks and install OpenClaw.
3. Choose manual onboarding, local gateway, loopback bind, and generate a token.
4. Register an OpenRouter account → add credits → create an API key → paste into OpenClaw to access models.
5. Create a Telegram bot with BotFather → paste bot token → set an allowlist (Telegram user IDs).
6. Complete onboarding and open the dashboard URL to test the initial conversation.
7. Run the agent security audit; ask the agent to apply recommended fixes; enable log redaction.
8. Add Brave (or Perplexity) API key to enable web search and test web fetch.
9. Create a daily automation (e.g., daily AI news); start small and iterate.
10. Use MD files to structure projects; build out files then spawn specialized agents as needed.

Limitations / cautions

• Don’t run OpenClaw on a machine with sensitive personal data unless you’ve hardened it.
• One‑click server images trade convenience for flexibility and can make later configuration harder.
• Cost/comparison numbers shown in subtitles may be rough — verify current model pricing before heavy usage.

Next video teased

• Building an agent team and a command center to coordinate multiple agents on projects.

Main speakers / sources

• Presenter / YouTuber (unnamed in subtitles) — demonstrates install and configuration on a Raspberry Pi.
• Services referenced: OpenClaw, OpenRouter, Telegram (BotFather), Brave Search (and Perplexity), OpenAI/Anthropic/Claude/Opus, open‑source models.