Running AI tools locally can be a practical choice when you want faster iteration, lower ongoing costs, or more control over data flow. This tutorial explains a clear, repeatable way to install OpenClaw with Ollama, verify everything is working, and fix common setup issues.
What you’re installing (in plain terms)
- Ollama: a local runtime that helps you download and run language models on your machine.
- OpenClaw: an app/tooling layer that connects to a local model runtime (like Ollama) so you can use models through a structured interface.
Prerequisites
- A supported OS (macOS, Windows, or Linux).
- Admin permissions to install software.
- Enough disk space for at least one model (often several GB).
- A terminal (Command Prompt/PowerShell on Windows, Terminal on macOS/Linux).
Step 1: Install Ollama
- Download and install Ollama from its official distribution for your OS.
- After installation, open a terminal and confirm it’s available by running a version or help command (for example, listing available commands).
Sanity check: can Ollama run a model?
Before adding OpenClaw, confirm Ollama can download and run at least one model locally:
- Use Ollama to pull a model (download it).
- Start a quick run session and send a simple prompt (e.g., “Write one sentence about local AI”).
If this works, your runtime is functioning and you can move on.
Step 2: Install OpenClaw
OpenClaw installation methods vary (package manager, installer, or container). Choose the one recommended by the project documentation. In most setups you’ll do one of the following:
- Install via package (common for developer-focused tools): download the release or install via a package manager.
- Run via container: use Docker/Podman if the project provides an image.
Keep it simple: install and confirm the app starts
- Complete the install method for your OS.
- Launch OpenClaw once to ensure it starts without errors.
Step 3: Connect OpenClaw to Ollama
This is the key step: OpenClaw needs to know where Ollama is running.
- Ensure Ollama is running in the background (some OS installs run it as a service; others require you to start it).
- In OpenClaw settings (or configuration file), select Ollama as the provider/runtime.
- Set the base URL/host to the local Ollama endpoint (commonly a localhost address).
- Pick a model that you already pulled in Ollama.
Verification test
Run a short prompt from inside OpenClaw. If you see a normal response (not an error about connection or missing model), the integration is complete.
Step 4: Install or select models (practical advice)
For a smooth first experience:
- Start with a smaller model if your machine is mid-range.
- Use a single model initially; add more after your pipeline is stable.
- Keep an eye on disk usage—models accumulate quickly.
Common problems and fixes
1) OpenClaw can’t connect to Ollama
- Check Ollama is running: restart the service/app and try again.
- Verify the host/port: confirm OpenClaw’s configured endpoint matches the local Ollama address.
- Firewall/VPN: temporarily disable or allow localhost traffic if something is interfering.
2) “Model not found” or empty model list
- Pull the model in Ollama first, then refresh OpenClaw.
- Ensure OpenClaw is pointing to the same Ollama instance where you downloaded the model.
3) Slow responses
- Switch to a smaller model.
- Close other heavy apps to free RAM/CPU.
- If available, enable hardware acceleration options recommended by your OS/runtime.
4) Installation conflicts or missing dependencies
- Re-run the installer with admin rights.
- On Linux, ensure required system packages are installed (varies by distro).
- If using containers, confirm Docker/Podman permissions and that the daemon is running.
Recommended next steps
- Create a repeatable workflow: document the model you use, the endpoint, and any environment variables.
- Build a simple prompt template: keep a “system” style instruction and a few reusable task prompts.
- Validate outputs: local AI is powerful, but you still need basic QA—especially for customer-facing text.
Once OpenClaw is talking to Ollama reliably, you have a stable base for experimenting with local AI features—without depending on a hosted API for every request.