How to set up OpenClaw with Ollama to run a local AI gateway

OpenClaw works with Ollama to create a local AI gateway for chat-based in­ter­ac­tions, agent workflows and tools. This means you can run large language models directly on your own system. In this guide, you’ll learn how to install OpenClaw, load a model, and set every­thing up, including the dashboard, messaging channels and skills (agent ca­pa­bil­i­ties).

Step 1: Set up your own server as the foun­da­tion

Running OpenClaw with Ollama on your own system comes with several clear ad­van­tages:

1. All model pro­cess­ing stays on your own in­fra­struc­ture. This is often the key reason, es­pe­cial­ly for data pro­tec­tion, con­fi­den­tial­i­ty, and com­pli­ance.
2. Setup is trans­par­ent. OpenClaw is open source, has clear doc­u­men­ta­tion and is fully con­fig­urable.
3. You keep full control over model selection, skills, services and startup behavior.
4. You can run the system without external model APIs, as long as you stick to local models and tools.
5. OpenClaw uses an agent-based design. It’s not just a chat interface, but a full working en­vi­ron­ment with skills, tools, a workspace, and a web interface.

On a dedicated server, you can set up an OpenClaw Docker in­stal­la­tion, which works seam­less­ly with Ollama.

Step 2: Check the re­quire­ments

Before you get started, make sure you have the following in place:

• A supported operating system: Linux, macOS, or Windows
• A working terminal or shell en­vi­ron­ment such as WSL2, SSH or Docker
• Internet access for the initial in­stal­la­tion
• Enough system resources: for larger models, plan for at least 8 to 16 GB RAM, ideally around 12 GB VRAM, and a CPU in the range of an Intel Core i5 or AMD Ryzen

OpenClaw also requires Node.js 22 or newer. If you use the official installer, you usually don’t need to install Node.js yourself. The installer sets up the runtime au­to­mat­i­cal­ly if needed.

Step 3: Install Ollama

Ollama is a tool for running LLMs locally on your system. OpenClaw connects to Ollama to handle chats, analysis and agent tasks. Ollama supports macOS, Windows, and Linux and au­to­mat­i­cal­ly provides a local API after in­stal­la­tion.

You can install it via the official Ollama website or directly through the terminal:

Step 4: Install OpenClaw

OpenClaw acts as the central execution layer in your setup. It receives requests from the web interface, loads the workspace with your selected skills. It also runs local tools such as shell commands or Python scripts and co­or­di­nates every­thing between the interface, agent logic and model.

To install OpenClaw on Windows, run the following command in the terminal:

iwr -useb https://openclaw.ai/install.ps1 | iex

On Linux or macOS:

curl -fsSL https://openclaw.ai/install.sh | bash

Then, start the on­board­ing process:

openclaw onboard --install-daemon

Step 5: On­board­ing

During on­board­ing, OpenClaw guides you through setting up your local AI en­vi­ron­ment. The process is in­ter­ac­tive and keeps things as simple as possible:

Close

As part of this, OpenClaw sets up a local gateway. This acts as an interface that other programs can use to com­mu­ni­cate with the AI. By default, the gateway runs only on your system, so it isn’t ac­ces­si­ble from external networks.

You’ll also set up your model during on­board­ing. For more complex agent workflows, a context window of around 64,000 AI tokens is helpful. Smaller context windows can cause the agent to lose track of longer con­ver­sa­tions or multi-step tasks. Common tool-capable models in Ollama include:

• qwen3-coder
• glm-4.7-flash
• gpt-oss:20b
• qwen3:8b

If your preferred model isn’t listed, you can also search for it manually. By default, OpenClaw au­to­mat­i­cal­ly downloads and sets up glm-4.7-flash.

At the end of on­board­ing, the gateway is con­fig­ured with a port, local binding, and an access token so that only au­tho­rized ap­pli­ca­tions can connect.

Step 6: Connect a messaging channel and set up web search

One of OpenClaw’s main features is its support for messaging services. This lets you send tasks directly from apps like WhatsApp, Telegram, Slack, Discord or iMessage to your local agent. You can set a channel up either during on­board­ing or later:

openclaw channels list 
openclaw channels add

The setup assistant walks you through au­then­ti­ca­tion for each platform:

• WhatsApp: Scan the QR code using the “Link devices” function
• Telegram: Enter your bot token (created via @BotFather)
• Slack / Discord: Au­then­ti­cate via OAuth in your browser
• iMessage: Available on macOS with an active iMessage login

Once connected, you can send tasks directly from the app. OpenClaw processes them locally using your Ollama model.

OpenClaw will also ask you to specify a web search service. For example, you can connect a Gemini sub­scrip­tion via API key to enable online search.

Step 7: Set up skills

Skills extend OpenClaw with custom, locally available ca­pa­bil­i­ties. Each skill is stored in its own folder and includes a SKILL.md file. This file combines metadata (YAML) with in­struc­tions (Markdown) that define how the agent behaves. OpenClaw au­to­mat­i­cal­ly loads skills from the workspace directory and detects changes if skill watching is enabled. Skills work par­tic­u­lar­ly well for recurring tasks, preferred workflows or project-specific knowledge.

During on­board­ing, you can select and configure several built-in skills, such as Google in­te­gra­tions, speech-to-text APIs, or sum­ma­riza­tion tools.

Close

Custom skills are typically created under ~/.openclaw/workspace/skills/. A good starting point is a skill that turns raw notes into struc­tured meeting minutes. It gives OpenClaw a clear task you can use in everyday work without needing complex workflows or tool in­te­gra­tions right away.

mkdir -p ~/.openclaw/workspace/skills/meeting-notes 
nano ~/.openclaw/workspace/skills/meeting-notes/SKILL.md

Then add the following:

---
name: meeting_notes
description: Creates clear, easy to follow meeting notes with tasks and open items from raw notes.
---
# Meeting Notes Skill
Use this skill when the user wants to turn messy meeting notes,
bullet points, or conversation transcripts into a clear summary.
Your task:
- Summarize the most important topics in a few sentences.
- Create a list of the decisions that were discussed.
- Clearly separate tasks, open questions, and next steps.
- Write in an objective, precise, and easy-to-read style.
- If information is missing or unclear, flag it clearly.
Desired output format:
1. Summary
2. Decisions
3. Tasks
4. Open questions
5. Next steps

Once saved, OpenClaw au­to­mat­i­cal­ly detects the new skill. If skill watching is enabled, updates to SKILL.md are applied without restart­ing the gateway.

Step 8: Activate hooks

Hooks are a flexible, event-driven au­toma­tion system. They let you trigger actions au­to­mat­i­cal­ly when certain events occur, such as when an agent runs a command. OpenClaw detects hooks au­to­mat­i­cal­ly when they’re placed in the correct di­rec­to­ries. You can manage them via the command line.

By default, these hooks are available during on­board­ing:

• session-memory: Au­to­mat­i­cal­ly saves session context to (~/.openclaw/workspace/memory/) when you run the /new command.
• bootstrap-extra-files: Loads ad­di­tion­al files into the workspace based on pre­de­fined paths or file patterns.
• command-logger: Logs all executed commands and saves them in ~/.openclaw/logs/commands.log.
• boot-md: Au­to­mat­i­cal­ly runs a BOOT.md file when the gateway starts, which can include startup in­struc­tions or ini­tial­iza­tion steps for the agent.

You can list all hooks using openclaw hooks list. You can also use openclaw hooks enable xy to enable in­di­vid­ual hooks.

Step 9: Start gateway

Once every­thing is installed and con­fig­ured, you can start the gateway:

openclaw gateway --port 18789

This launches the local gateway and makes the dashboard available at: `http://127.0.0.1:18789``.

To check if every­thing is running correctly, use:

openclaw gateway status

Look for these status messages: “Runtime: running” and “RPC probe: ok”. These indicate that the system is working as expected.

Step 10: Test dashboard

Most in­ter­ac­tions with the bot happen through the messaging services you’ve set up, so the built-in dashboard isn’t something you need for day-to-day use. That said, it’s still worth taking a quick look during setup. It gives you a clear overview of whether the gateway is running, whether the model was detected correctly, and whether your con­fig­u­ra­tion is valid. If you’re running OpenClaw on a desktop system, you can also use the dashboard as an oc­ca­sion­al di­ag­nos­tic tool.

Close

You can open the dashboard directly in your browser or via the CLI:

openclaw dashboard

You can also open the web interface using the gateway address `http://127.0.0.1:18789``. If you’re prompted to au­then­ti­cate the first time you access it, you can retrieve the gateway token using:

openclaw config get gateway.auth.token

Then enter this token in the control UI. Local con­nec­tions via 127.0.0.1 are usually accepted au­to­mat­i­cal­ly. If you’re accessing the dashboard from another device, you may need to approve it first:

openclaw devices list 
openclaw devices approve <requestId>

This marks the new browser or device as trusted. That extra approval step is important because the dashboard isn’t just a chat interface. It’s an ad­min­is­tra­tive interface with access to skills, con­fig­u­ra­tion, and execution per­mis­sions.

Step 11: Customize OpenClaw

At this point, you’ve got the basics up and running. The gateway is running locally, the model is connected and you can interact with the agent through messaging services or the dashboard. Even with this setup, you can already start using OpenClaw to analyze messages, summarize in­for­ma­tion, or automate simple tasks.

That said, OpenClaw is built to be modular, and there’s a lot more you can do beyond this initial setup. Some common ways to expand it include:

• Develop ad­di­tion­al skills: Set up recurring tasks like meeting summaries, code analysis, or document struc­tur­ing as dedicated skills the agent can use anytime.
• Build au­toma­tions with hooks: Use events like new sessions, command ex­e­cu­tions, or agent starts to trigger actions au­to­mat­i­cal­ly.
• Integrate local tools: OpenClaw can run shell commands, Python scripts, or your own tools, turning it into a kind of local au­toma­tion platform.
• Combine multiple models: Alongside your main model, you can add fallback or spe­cial­ized models for specific tasks.
• Connect more messaging services: Add channels like Slack, Discord, or Telegram so multiple devices or teams can work with the agent.
• Expand the workspace and knowledge base: Store documents, project files, or notes so the agent can use them as context.
• Set up remote access: With tools like Tailscale or reverse proxies, you can securely access the dashboard from other devices.
• Look for in­spi­ra­tion: The “What People Are Building” page shows what other users are building with OpenClaw.

What you’ve set up here is really just the starting point. Once every­thing is running smoothly, you can gradually build out more advanced workflows with ad­di­tion­al skills, au­toma­tions and in­te­gra­tions. Es­pe­cial­ly when combined with local models through Ollama, OpenClaw becomes a flexible, fully self-hosted AI workspace you can shape around your own needs.