Turning Your Local Model Into an Always-On Bot with OpenClaw

What OpenClaw actually is

Everything else in this pillar gets you a model you can talk to in a browser tab. That's genuinely useful, but it only works while you're sitting at that machine with that tab open. OpenClaw is a different piece of the stack: a self-hosted agent framework that runs as a background service and gives your model a persistent presence in Discord or Telegram - plus memory files it reads on every turn, so it can actually remember things about you and your setup across conversations.

Pointed at a cloud provider, it works like any hosted AI bot. Pointed at your local Ollama install, it's the same private, no-monthly-bill setup this whole site is about, just reachable from your phone instead of tied to one machine.

Installing OpenClaw

The official installer script handles Node.js and launches the onboarding wizard in one step:

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

It detects your OS, installs Node.js if it's missing, installs OpenClaw, and starts onboarding automatically. During the wizard, when you reach the model/auth provider prompt, that's where you'll point it at Ollama instead of a cloud provider - covered next.

Connecting it to Ollama

Before touching OpenClaw's config, confirm Ollama itself is actually running and reachable:

curl http://localhost:11434/api/tags
# Should return JSON listing your pulled models
# "connection refused" means Ollama isn't running - start it with: ollama serve

The simplest way to wire the two together is a single environment variable - OpenClaw then auto-discovers every model you have installed:

export OLLAMA_API_KEY="ollama-local"
echo 'export OLLAMA_API_KEY="ollama-local"' >> ~/.bashrc
source ~/.bashrc

Restart OpenClaw after setting this, and it will query Ollama directly for your model list, read their context windows, and set every cost to $0 automatically.

What actually goes wrong

Tool calls come out as plain text instead of actually running

This is almost always a URL mistake: if you're hand-writing a config block, the Ollama base URL must not have /v1 on the end. Adding it switches to Ollama's OpenAI-compatible endpoint, which breaks tool calling entirely - your bot will print raw tool JSON as a chat message instead of executing it. Use the plain native URL, http://127.0.0.1:11434, with "api": "ollama" set explicitly.

OpenClaw doesn't see any models

Confirm Ollama is actually reachable first with the same curl http://localhost:11434/api/tags check from above, then confirm OpenClaw's side:

openclaw models list
openclaw models status

Responses feel slow with more than one person messaging it

This is expected, not a misconfiguration - Ollama processes one request at a time, so on CPU-only hardware, a second person's message waits behind the first. Fine for a personal bot or a small group; a smaller model as the primary (with a larger one as fallback for when quality matters more) keeps things feeling responsive.

Common questions