OpenClaw is an incredible open-source AI agent framework. Downloading and installing it doesn't cost a dime. But the moment you actually start using it, prepare to watch your tokens burn at an alarming rate.
The costs of OpenClaw don't just come from the core model's replies. They accumulate from web reading, memory retrieval, summarization, tool calling, and all the workspace files and bootstrap configurations crammed into the system prompt. Once your context window gets long, your monthly bill will hit you like a truck.
Running OpenClaw with Claude 3.5 Sonnet, a single month of accumulating 10 million input tokens and 10 million output tokens can easily cost you nearly a hundred dollars. If you treat it as a 24/7 executing Agent running complex tasks on high-tier models, burning through thousands of dollars a month is not an exaggeration. Case in point: OpenRouter recently saw its processed token volume skyrocket from 6.4 trillion to 13 trillion a week.
You wanted AI to work for you, but you ended up handing your entire paycheck over to the AI.
Since cloud token expenses are brutally high, running the stack locally is the obvious choice. This is where pairing OpenClaw with Ollama saves the day.
Ollama handles running open-source models like Llama, Mistral, or DeepSeek directly on your local machine, completely zeroing out your API costs. Aside from your graphics card fans spinning a bit faster, there are no downsides. Plus, all your private data and code processing stay strictly on your local hardware, requiring no cloud uploads. This keeps costs predictable while guaranteeing absolute privacy.
In Practice: Combining OpenClaw with Ollama Locally
OpenClaw is a framework built on Node.js, and it strictly requires a Node.js 22 or higher environment to run.
This is where you can use ServBay to deploy your Node.js environment.
As a local web development environment management tool, ServBay handles multiple versions of Node.js effortlessly. Through its clean graphical interface, you can quickly switch to a Node.js 22 environment. It allows you to install Node.js environment with one click, completely bypassing the headache of manually configuring environment variables or fighting version conflicts.
Once your environment is ready, deploying OpenClaw takes just a couple of simple commands:
curl -fsSL https://molt.bot/install.sh | bash
openclaw onboard --install-daemon
Using ServBay, you can also download and install Ollama with a single click. From there, simply select the large language model you want from the left-hand menu and download it.
OpenClaw doesn't have the ability to "think" on its own out of the box; it needs to connect to Ollama using the following command:
ollama launch openclaw
This command configures OpenClaw to use the models provided by your local Ollama instance.
Security First: Git Safety Nets and Permission Controls
When an AI gains operational privileges over your system, the security risks escalate exponentially. We've all seen the horror stories of OpenClaw accidentally deleting entire email inboxes.
An agent with execution permissions can destroy a system if it misunderstands an instruction. To counter these potential disasters, we must implement strict defense mechanisms.
Git as a Safety Net
OpenClaw highly recommends placing your entire workspace (including configuration files and memory logs) under Git version control.
git init
git add AGENTS.md SOUL.md memory/
git commit -m "Initialize agent workspace"
If the agent installs a broken skill during a task or makes erratic changes to configuration files, you can simply use git revert to roll the system state back to a safe point in time. This version-controlled evolution makes the AI's behavior transparent and entirely reversible.
Permission Limits and Sandbox Mode
An agent's capabilities come entirely from its skill system. To prevent third-party skills from injecting malicious code, you should manually audit the source code and confirm the exact commands it executes before installing anything. Furthermore, for agents handling highly complex tasks, it is strongly recommended to run them in isolated environments, such as Virtual Machines or Docker containers.
Authentication and Private Access
The Gateway service should never be exposed directly to the public internet. The secure approach is to enable gateway authentication and run openclaw doctor to diagnose potential risks. When accessing it remotely, pair it with a VPN or an intranet tunneling tool to ensure that only authorized users can send commands to your agent.
Conclusion
OpenClaw is fantastic. It's a great toy to play around with, but if you want it to act as a reliable, 24-hour employee, the cloud costs and operational risks are still incredibly high. Running it locally with Ollama and sandboxing its environment is the only sustainable way forward.
Top comments (0)