What if your AI assistant never called home? No OpenAI, no Anthropic, no external API, just local models running on your own hardware, on your own network, completely under your control.
You can run OpenClaw that way today. OpenClaw’s 2026.4.27 release quietly added a config flag, models.pricing.enabled: false, that disables all external pricing and model-discovery fetches at startup. Pair it with a local model runner like Ollama or LM Studio and you get zero outbound internet dependencies. Fully offline. Fully yours. The rest of this guide is about which setup fits you: weigh the criteria, pick an approach, then wire it up.
What Matters Before You Go Offline
Going offline is a trade between a few things. Weigh these before you commit:
Privacy. Every prompt you send to a cloud API leaves your machine. Even with generous privacy policies, there’s inherent risk in routing personal, proprietary, or sensitive queries through third-party infrastructure. Local models mean your data never leaves your network.
Cost. Per-token fees add up, especially for heavy users, automated workflows, or anyone running OpenClaw as part of a larger pipeline. Once you’ve downloaded a local model, inference is essentially free. Your electricity bill might tick up slightly; your API bill goes to zero.
Model quality. The best local models are impressive, but there’s still a gap compared to GPT-4o or Claude 3.5 Sonnet on complex reasoning tasks. For everyday work like summarizing, drafting, answering questions, and light coding, it’s a non-issue. For deep analysis or challenging multi-step reasoning, you’ll feel the difference.
Hardware. You don’t need a beast of a machine, but you do need enough RAM to run a model comfortably:
- Minimum: 8GB RAM (runs Phi-3 Mini or small 4-bit quantized models)
- Recommended: 16GB RAM (comfortable with Llama 3.1 8B, Mistral 7B)
- Better: 32GB+ RAM or a GPU (opens up 13B+ models, faster inference)
- Raspberry Pi: Doable with Phi-3 Mini, slow but functional
Apple Silicon Macs are particularly well-suited here. The unified memory architecture means models load fast and inference is snappy even on base M-series chips.
Ease. Ollama is the path of least resistance. It handles model downloads, quantization, and serving behind a clean CLI and an OpenAI-compatible API endpoint. LM Studio offers the same with a GUI.
There’s also a practical reason some people have no choice: restricted environments. Corporate networks with egress filtering, government or compliance-sensitive setups, homelab machines on isolated VLANs, Raspberry Pi builds in the field, sometimes you literally can’t reach the internet, or you’re not supposed to. This setup handles all of that.
Three Ways to Run OpenClaw Offline
Depending on how strict you need to be, there are three ways to set this up:
1. Local model, still online. Run Ollama or LM Studio on your machine and point OpenClaw at it. Inference never leaves your box, but OpenClaw still makes its usual startup calls to external services unless you disable them.
2. Fully air-gapped. Run the local model, set models.pricing.enabled: false, and disable the heartbeat external check. Nothing reaches the internet, not even at startup.
3. Hybrid. Keep inference local but leave connectivity for the integrations that genuinely need it, like Discord bots, webhook listeners, and SMTP email checks. Here “offline” means the AI inference layer only; the networking features stay separate.
Head-to-Head: Local vs Air-Gapped vs Hybrid
Local only is the fastest to set up and fine if your main goal is keeping prompts off third-party APIs. But it still phones home at startup, so it isn’t truly air-gapped until you add the pricing flag.
Fully air-gapped is the strongest on privacy and reliability. Once the flag and heartbeat are off, OpenClaw points entirely inward and nothing sneaks out at startup. It costs you a little extra config, nothing more.
Hybrid is the practical middle ground if you run bots, webhooks, or email integrations that have to reach the network. Your prompts stay local; only the integration traffic goes out.
The takeaway: if privacy or compliance is the driver, go air-gapped. If you just want local inference with less fuss, local plus the pricing flag gets you most of the way. If you need connected integrations, go hybrid.
Set Up Your Local Model
You’ll need:
- OpenClaw: version 2026.4.27 or later (for the
models.pricing.enabledflag) - Ollama: the easiest local model runner, highly recommended
- A local model: Llama 3, Mistral 7B, Phi-3, or similar (more on this below)
Alternatively, LM Studio works if you prefer a GUI. The OpenClaw config will be the same either way.
Install Ollama:
# macOS
brew install ollama
# Linux
curl -fsSL https://ollama.com/install.sh | shOr download the installer from ollama.com.
Pull a model:
ollama pull llama3.1This downloads Llama 3.1 8B by default, about 4.7GB. For something lighter:
ollama pull phi3 # ~2.3GB, great for low-RAM setups
ollama pull mistral # ~4.1GB, solid general-purpose chatStart the server (if it isn’t already running):
ollama serveLM Studio alternative: If you prefer a graphical interface, LM Studio lets you browse, download, and run models through a desktop app. It also exposes a local server on port 1234 with OpenAI-compatible endpoints. The OpenClaw config step below works the same way, just change the port and base URL accordingly.
Point OpenClaw at Your Local Provider
OpenClaw supports OpenAI-compatible APIs out of the box. Ollama speaks that protocol, so wiring them together is straightforward.
In your OpenClaw config (typically ~/.openclaw/config.yaml or via openclaw config), add or update your providers section:
providers:
- name: ollama-local
type: openai-compatible
baseUrl: http://localhost:11434/v1
models:
- llama3.1
- mistral
- phi3
defaultModel: ollama-local/llama3.1If you’re using LM Studio instead:
providers:
- name: lmstudio-local
type: openai-compatible
baseUrl: http://localhost:1234/v1
models:
- your-loaded-model-name
defaultModel: lmstudio-local/your-loaded-model-nameAt this point, OpenClaw will route completions to your local model. But it’ll still try to fetch external pricing and model discovery data on startup. That’s where the next step comes in.
Stop the Startup Calls (Go Air-Gapped)
The key addition in OpenClaw 2026.4.27 is this config flag:
models:
pricing:
enabled: falseWhen models.pricing.enabled is set to false, OpenClaw skips the startup calls it normally makes to external services (OpenRouter, LiteLLM, and similar) to fetch model pricing data and discover available models. These calls happen even if you have no cloud providers configured, and they’re the main thing that stops a “local-only” setup from being truly offline.
Add this to your config and those fetches stop entirely.
If you want to go fully air-gapped, also disable the heartbeat external check:
heartbeat:
enabled: falseOr if you want to keep heartbeats but ensure they don’t make external HTTP calls, review your heartbeat.md and remove any checks that reach outside the LAN.
Your final config additions for a fully offline setup should look something like this:
providers:
- name: ollama-local
type: openai-compatible
baseUrl: http://localhost:11434/v1
models:
- llama3.1
defaultModel: ollama-local/llama3.1
models:
pricing:
enabled: false
heartbeat:
enabled: falseSave the config, then restart OpenClaw.
Verify You’re Actually Offline
Time to verify you’ve actually gone offline.
Disconnect from the internet. On a Mac, turn off Wi-Fi and unplug Ethernet. On Linux, sudo ip link set eth0 down or just physically disconnect.
Send a message through OpenClaw. Ask it something basic, “What’s the capital of France?” or “Write a haiku about databases.” You should get a response from your local model.
If it works: you’re done. Full AI, zero cloud.
Once it’s running, the right model depends on your hardware and workload:
| Model | Size | Best For | Min RAM |
|---|---|---|---|
| Llama 3.1 8B | ~4.7GB | General chat, Q&A, writing | 8GB |
| Mistral 7B | ~4.1GB | Chat, reasoning, instructions | 8GB |
| Phi-3 Mini | ~2.3GB | Lightweight, Raspberry Pi, fast | 4GB |
| CodeLlama 7B | ~3.8GB | Code generation, debugging | 8GB |
| Llama 3.1 70B | ~40GB | High quality, near-GPT-4 | 64GB |
For most people on a modern laptop or desktop with 16GB RAM, Llama 3.1 8B is the sweet spot: fast enough to feel snappy, capable enough to be genuinely useful.
If you’re on a Raspberry Pi 4 or 5 with 8GB RAM, Phi-3 Mini is where you want to start. It’s impressively capable for its size and can run on modest hardware.
If you’re doing code-heavy work, CodeLlama is worth the download even if you also have a general-purpose model running.
The Tradeoffs, Honestly
Local inference isn’t a free lunch. Here’s what you give up:
Speed. Local inference on CPU is slower than cloud APIs, sometimes significantly. A cloud API response arrives in a second or two; a 7B model on CPU might take 5 to 20 seconds depending on the prompt length and your hardware. GPU acceleration helps a lot if you have it.
Image generation. If you use OpenClaw’s image generation features, going fully offline requires a local diffusion model setup (Stable Diffusion, ComfyUI, etc.). That’s a heavier lift and beyond the scope of this guide.
Some channels expect connectivity. Certain OpenClaw integrations, Discord bots, webhook listeners, SMTP email checks, obviously need network access to function. “Offline” here means the AI inference layer; the networking features are a separate concern.
The Verdict
The fully offline OpenClaw setup is genuinely useful, not just a proof of concept. Once it’s running, it works. It’s fast enough for real work on decent hardware, the models are good enough for most tasks, and the privacy and reliability benefits are real.
Who’s this for? Privacy-conscious users who don’t want prompts leaving their machine. Developers building on restricted corporate networks. Homelab and Raspberry Pi enthusiasts who want AI as a local service. Anyone who’s been burned by API outages at exactly the wrong moment. And anyone who just likes the idea of owning their tools outright.
The models.pricing.enabled: false flag is a small thing. But it’s the piece that closes the loop: it means you can point OpenClaw entirely inward and trust that nothing is sneaking out to the internet at startup.
Full AI, full control, zero cloud dependency. That’s the setup.
Running OpenClaw in another interesting configuration? Drop it in the comments.
Not Ready to Self-Host? Try ClawTrust
Building agents locally is powerful, but managing infrastructure, security, and scaling takes effort. ClawTrust provides managed OpenClaw hosting — your agents run securely in the cloud with automatic scaling, no DevOps overhead, and 24/7 monitoring.
This post contains affiliate links. If you buy through them, I may earn a small commission at no extra cost to you.
About The Author
This is exactly what I have been waiting for. Running anything cloud-dependent on a corporate network is a compliance nightmare. Going to test this with Ollama this weekend.
Let us know how it goes. I set it up with LM Studio and the only hiccup was pointing the model path correctly in the config. Once that was sorted it just worked.
Great question. For Run OpenClaw Completely Offline Air, I would start with the simplest workable setup and only add complexity if you actually need it. That keeps things easier to troubleshoot and usually gets you to a better result faster.
Thanks for reading, Neil Forsythe. For Run OpenClaw Completely Offline Air, I would start with the simplest workable setup and only add complexity if you actually need it. That keeps things easier to troubleshoot and usually gets you to a better result faster.
Does this work with Mistral 7B or is it better to go with Llama 3 for general assistant tasks? Trying to figure out the right model before I commit to downloading anything.
Great question. For Run OpenClaw Completely Offline Air, I would start with the simplest workable setup and only add complexity if you actually need it. That keeps things easier to troubleshoot and usually gets you to a better result faster.
The cost angle alone sold me. I was burning through API credits faster than I expected. Having a free inference layer for routine tasks makes a lot of sense.
Thanks for reading, Caleb Marsh. On Run OpenClaw Completely Offline Air, I would start with the basics and then tighten things up once you see how it behaves in practice. If you are already testing it, the main thing is to keep the setup simple enough to debug quickly.