OpenClaw combines a self-hosted assistant, coding agent, messaging bot, and scheduler. Several forks use the same basic idea with different resource requirements. This guide covers the initial setup, model routing, useful skills, and the reasons to choose a fork.
Treat it as the hub. Each section is a short orientation that links to a full walkthrough, so come back here whenever you’re deciding what to do next. If you don’t yet know what OpenClaw is at all, read the explainer first.
The five things that make up OpenClaw
Hold these five pieces and the rest falls into place:
OpenClaw at a glance
| Piece | Role |
|---|---|
| Model provider | The brain, DeepSeek, GLM, Kimi, Qwen, or a local model |
| Channels | How you talk to it, terminal, Telegram, Discord, Slack |
| Skills | What it can do, including the coding agent |
| Scheduler | When it acts, unattended cron-style jobs |
| The daemon | The background process tying it all together |
You bring the model; OpenClaw supplies the rest. That’s the whole design, and it’s why this guide is organized around the model first, then everything you can do with it.
Step 1: get it running
Start with installation, because nothing else matters until the daemon runs. On Windows the clean path is WSL2, OpenClaw expects a Unix-style environment, and running directly in PowerShell is the rougher road. The Windows and WSL install guide walks through it; if you’re on a fresh machine, the general WSL2 setup gets the environment ready first.
Step 2: pick a model
This is where you control cost and privacy. OpenClaw is model-agnostic, so the choice is yours:
- DeepSeek uses pay-per-token billing and publishes OpenAI-compatible routes.
- GLM has subscription tiers for heavier daily use.
- A local model through Ollama avoids an API bill and keeps prompts on your machine.
- Kimi, Qwen, and MiniMax are also available through compatible routes. Check OpenClaw’s provider documentation and the live model price before configuring one.
DeepSeek is a hosted pay-per-token option, while Ollama runs a model locally. Compare current hosted rates and local hardware requirements in the AI coding API comparison.
Step 3: use the skills
Skills are what turn a model into an assistant that acts. The coding agent is the headline one, it reads issues, writes fixes, and opens PRs as a background process, covered in the coding agent tutorial. Beyond that, browse the project’s own skill docs and install only the skills that match work you actually want automated.
Step 4: decide between OpenClaw and a fork
OpenClaw is the most feature-complete option, and also the heaviest. The fork wave produced lighter alternatives, each with a clear trade:
OpenClaw and its main forks
| Project | The pitch |
|---|---|
| OpenClaw | Most features, most channels, the full experience |
| nanobot | Python, minimal, easy to read and audit |
| ZeroClaw | Rust implementation designed for smaller hosts and VPSes |
| PicoClaw | Go single binary, runs on a Raspberry Pi |
| NanoClaw | TypeScript, container-isolated for safety |
If you want everything, stay on OpenClaw. If you want light, auditable, tiny-hardware, or container-isolated, a fork or rewrite may fit better. The full comparison, including source links for nanobot, ZeroClaw, PicoClaw, and NanoClaw, is in OpenClaw vs its forks.
Step 5: keep cost and security in check
Two things to get right before you lean on it daily.
An always-on assistant can make repeated API calls overnight. Use conservative schedules, set provider budgets, and review usage. DeepSeek, GLM, and Ollama use different billing and hosting models.
Security. OpenClaw runs with real access to files, commands, and your messaging accounts. Scope it, keep keys in environment variables, and think before pointing it at anything sensitive.
A sensible first run
If you want a concrete starting path rather than the full menu:
Your first OpenClaw setup
- Install on Windows via WSL2 (or natively on macOS/Linux)
- Choose a hosted provider or a local Ollama model
- Talk to it from the terminal first, add Telegram/Discord later
- Enable one skill you'll actually use, the coding agent is a good start
- Scope permissions and put your API key in an environment variable
- Only add scheduled, unattended jobs once you trust the setup
Verify the result
OpenClaw has five main pieces: the model, channels, skills, scheduler, and daemon. Choose a model that fits your budget and privacy needs, enable only the skills you use, and decide whether the main project or a lighter fork fits the host. Start with a small WSL2 setup and add capabilities after the basic workflow is stable.
Every section above links to a full guide. If you’re starting cold, go install it, then point it at DeepSeek. If you’re comparing options first, read OpenClaw vs the forks and OpenClaw vs Claude Code.
