Summary of "OpenClaw Full Course: Setup, Skills, Voice, Memory & More"
Purpose
Step-by-step, in-depth tutorial to install, configure and make an OpenClaw (Clawbot / Claudebot) instance productive and secure. Emphasis on practical best practices learned from many hours of personal configuration.
Security & deployment
- Treat OpenClaw as an untrusted virtual assistant. Isolate it from sensitive accounts (main computer, crypto keys, primary email).
- Use separate service accounts and limited-permission tokens for integrations.
- Prefer running OpenClaw on a VPS rather than local hardware. VPS advantages:
- automatic backups, data-center physical security,
- easy on/off and low cost.
- Referenced a longer (~50-minute) security/hardening video for advanced hardening steps.
Best practice: assume the agent can be compromised and minimize its access to sensitive resources.
Quick install (Hostinger one‑click)
Recommended fastest route: Hostinger KVM2 one‑click deploy. (Presenter referenced a discount code.)
Typical flow:
- Choose plan and location.
- Paste one LLM API key (Anthropic Opus in the transcript).
- Deploy — Hostinger runs OpenClaw in a Docker container.
- After deploy: copy the gateway token from the environment, open the OpenClaw gateway web UI and log in with the token.
Server access and Docker
Options to interact with the VPS/container:
- Use the host web terminal.
- SSH to the VPS directly.
- SSH from VS Code or Cursor (remote).
Docker notes:
- OpenClaw runs inside a container.
- To run CLI inside the container: list containers (docker ps) then exec into the container (docker exec -it /bin/bash) and run openclaw commands.
- Use gateway logs (filter by error/fatal) to debug common problems (missing API keys, insufficient billing credits, etc.).
LLM model strategy & cost control
- Use a high-quality model (Anthropic Opus, e.g., Opus 4.6) for planning and high-leverage tasks.
- Use a cheaper fallback model (OpenAI Codex / ChatGPT subscription / “codeex”) for routine tasks and code execution to save cost.
- Configure automatic model routing:
- Set Opus as default for planning.
- Delegate coding and routine tasks to the cheaper model.
- Example approach: create a /model command to switch models and save routing rules; have tasks report which model handled them.
- Billing tips: add API credits, set monthly spend limits, enable spend notifications.
Integrations — Telegram
- Create a Telegram bot via BotFather and copy the bot token.
- Pair the bot with OpenClaw using the gateway pairing command.
- Promote the bot to admin in group chats and configure permissions.
- Create separate Telegram groups for scoped topics (e.g., “startup ideas”, “accounting”) so the bot’s memory/history and behavior can be channel-scoped.
- Configure rules so the bot restricts responses by channel topic and replies only when required.
Voice / speech-to-text
- Enable the audio transcription skill. Supported providers include Deepgram, OpenAI Whisper, etc.
- If a provider key is not required, OpenClaw may fall back to a default provider.
- If transcription fails, install Whisper locally or choose another provider.
- Once enabled, sending audio messages to Telegram will allow OpenClaw to transcribe them (voice mode).
Skills system (what a skill is, ClawHub, creating skills)
- Definition: a skill is a folder containing a skill.md (markdown) plus optional scripts (for example, Python files) that the agent can run.
- ClawHub provides pre-built skills. Caution: review and vet skills before enabling them.
- OpenClaw can write its own skills on demand. Example workflow:
- Create a “hello world in 5 languages” skill → folder + skill.md generated → registers a command (e.g., /skill hello-world-languages).
- Skill installation: enable from the Gateway and install dependencies (some skills require extra setup).
Coding agent & GitHub integration
- Enable the coding agent skill so the bot can generate code and scripts.
- Create a separate GitHub account for the bot and connect via GitHub CLI / OAuth.
- Configure git identity inside the container and allow OpenClaw to auto-commit code for auditing and backups.
- Use the coding agent + GitHub to let the bot create, commit, and manage its own repositories.
Editing workspace and files
- Use VS Code or Cursor remote SSH to browse the container’s OpenClaw data/workspace.
- Important locations to review:
- workspace/skills
- memory files
- config files
- scripts
- Reading skill.md and related scripts helps you understand and audit the bot’s capabilities.
Memory architecture & enhancements
Types of memory:
- memory.md: persistent long-term memory.
- Daily/session memory files: short-term per-day logs.
Recommended settings and features:
- Enable compaction and memory flush so important context is written to memory before history gets compacted.
- Enable session memory search so recent session transcripts are searchable.
- Vectorized memory search: enable a vector memory backend (transcript mentioned a “QMD” backend). Vector search greatly improves recall and retrieval quality.
- Follow OpenClaw documentation for prerequisites and setup of vector memory backends.
Identity, user, soul, tools, heartbeat
- user.md and identity.md: store human-specific and bot-specific long-term data. Best practice: have the bot interview you to populate these files rather than editing them manually.
- soul.md: bot persona, core beliefs, behavior boundaries and “vibe” — edit to adjust tone or constraints.
- tools.md: device and tool preferences used by skills.
- heartbeat.md:
- Defines periodic wake-up tasks (default/recommended interval often ~30 minutes, configurable).
- On each heartbeat the bot can run review jobs, spawn subagents, check email, or perform maintenance.
- Crons: scheduled one-off or recurring jobs for reminders and automation.
Subagents (background agents)
- Subagents are lightweight background agents spun up to do parallel work.
- Many subagents can be created to split tasks into parallel subtasks.
- They are visible in the Gateway sessions area and useful for parallelization.
Practical recommendations & tips
- Always check logs and billing when something stops working (common causes: missing API key or low credits).
- Start with a secure baseline; using the one-click VPS route speeds setup — harden later as needed.
- Keep sensitive data out of the agent: separate accounts and limited-permission tokens.
- Don’t enable random ClawHub skills without reviewing their contents.
- Balance quality vs. cost with a mix of high-quality and cheaper models.
- Use GitHub integration to automatically store code written by the agent for review and version control.
Tools / integrations mentioned
- Hostinger (VPS one‑click deployment)
- Docker (container runtime)
- Anthropic (Opus models)
- OpenAI (Codex / GPT)
- Telegram (BotFather, groups, tokens)
- Whisper / Deepgram (STT providers)
- ClawHub (pre-built skill marketplace)
- GitHub (CLI, OAuth, separate bot account)
- VS Code or Cursor (remote SSH editor)
- QMD / vector memory backend (for vectorized memory search)
- OpenClaw gateway (web UI), logs, sessions, skills dashboard
Referenced resources / further guides
- Longer (~50-minute) hardening/security video (shown in the original presentation).
- OpenClaw documentation and examples (memory & vector backend instructions).
- ClawHub for skill repositories.
Main speaker / sources
- Presenter: Tech With Tim (Tim). Coupon code referenced: “techwithtim”.
- Technologies and providers discussed: OpenClaw, Hostinger, Anthropic (Opus), OpenAI (Codex/ChatGPT), Telegram, Docker, Whisper/Deepgram, GitHub, ClawHub, VS Code/Cursor.
Category
Technology
Share this summary
Is the summary off?
If you think the summary is inaccurate, you can reprocess it with the latest model.
Preparing reprocess...