How to run OpenClaw with Ollama, what to pick during setup, why a small local model handles the agentic loop poorly, and where privacy actually stops.

OpenClaw can work with Ollama, but the phrase "run it locally" hides two different tasks. You can install the agent itself on your computer and keep using a cloud model. Or you can keep both the agent and the model local. In the second case, data is better controlled, but the hardware and model-quality requirements go up noticeably.

What OpenClaw is

OpenClaw is a personal agent with a gateway process, message channels, tools, and scheduled tasks. You can link it to messengers and leave it running as a service, not just open it for a single session.

That's what sets OpenClaw apart from a regular local chat. The model doesn't just answer — it plans actions, calls tools, and continues the loop based on results. The general pattern for an app like this is covered in the article on the local AI agent.

How to run OpenClaw with Ollama

Per the current OpenClaw guide, the easiest starting point is the openclaw onboard wizard:

Install and start Ollama.
Download a suitable model in Ollama.
Run openclaw onboard and pick local mode.
Point it at Ollama and the local server address, if the wizard doesn't find it automatically.
Check the catalog with openclaw models list --provider ollama.

For a regular single-computer install, OpenClaw looks for Ollama at http://127.0.0.1:11434. In manual configuration, don't add /v1 to the address: OpenClaw uses Ollama's native API to pass tool calls more reliably.

◆

Verify the model name through ollama list or OpenClaw's own catalog. A similar but inexact name causes a provider error, not an automatic fallback to a different model.

Why a chat model might not be a good fit

An agentic loop is heavier than a regular conversation. System instructions, tool descriptions, action history, and command results all land in the context. The model has to not just write an answer, but pick a tool, form correct arguments, notice an error, and continue.

OpenClaw's official page on local models warns directly: small, heavily quantized models hold context worse and are less protected against prompt injection. A comfortable autonomous loop can require several powerful machines; a single 24 GB GPU is better suited to lighter requests and tolerable latency.

That doesn't mean nothing will run at home. It means you should scope down the tools first and test a short scenario, rather than immediately handing the model email, a browser, and the terminal. How GPU memory relates to model size is described in the article on GPUs for AI models.

Where "local" stops

A local model in Ollama keeps its requests on your machine. But a connected Telegram, email, GitHub, or web search remain external services. A message passes through their servers, and a tool can send part of the context outward.

The same applies to a remote MCP server. Having the agent's client installed locally doesn't make a cloud integration private. Check, channel by channel, what data it receives, where it stores tokens, and which actions are allowed. The basic permission model is explained in the security section of the article on connecting an MCP server.

OpenClaw or Doka

OpenClaw is useful if you need an always-running gateway, message channels, and flexible server-side configuration. You pay for that with time spent on installation, model choice, permissions, and process maintenance.

Doka is built for a different start: install the desktop app, download a local model from the catalog, and work with files, documents, and the terminal in a chosen folder. For messengers and external services, you can add MCP, but Doka doesn't turn them into local services or promise to control any arbitrary window on the computer.

A practical first test

Don't start with a full day of autonomous work. Give the agent a local folder with no secrets and a 10–15 minute task: pull a summary from a few files, run a safe command, and save the result. Watch whether it asks for confirmation, whether it loses the goal, and what ends up in the log.

If you want a local agent with no gateway or Ollama setup, download Doka. If messengers, a remote server, and your own infrastructure matter, OpenClaw gives you more control — along with more responsibility for the configuration.

OpenClaw Locally: Installing With Ollama and Hardware Requirements