OpenClaw Runs on a Local LLM via Ollama or LM Studio

Yes: OpenClaw runs on a local LLM, and its docs support Ollama, LM Studio, an OpenClaw-managed llama.cpp server, and any server that speaks the OpenAI chat API. What it takes is a model that handles tool calls, a large context window, and tighter permissions than you would use with a hosted model.

OpenClaw is the open-source personal agent that connects chat apps such as WhatsApp, Telegram and Signal to an AI agent running on your own machine. It is MIT-licensed and run by the OpenClaw Foundation, and its latest release when we read the docs on 14 October was v2026.10.3. Its home page says you need “an API key from your chosen provider”, but a local server can stand in for that provider.

Your local options

OpenClaw’s Local models page lists these backends. The “use when” wording is the docs’ own.

Backend	How OpenClaw connects	OpenClaw’s docs say use it when
Ollama	Ollama’s native API at http://host:11434, not /v1	You want a CLI workflow, a model library and a background service
LM Studio	Its local server at http://127.0.0.1:1234/v1, Responses API	First-time local setup, with a GUI loader
llama.cpp	A server OpenClaw installs and manages, picked in onboarding	You want hardware-aware model selection and verified downloads
MLX, vLLM, SGLang	A custom provider pointed at their OpenAI-style endpoint	You want high-throughput self-hosted serving
ds4	OpenAI-compatible tool calls	You want DeepSeek V4 Flash locally on a Mac (Metal)
llmman	OCI-registry model pulls, with llama.cpp, vLLM or MLX engines	You want hybrid local and hosted routing
LiteLLM or another proxy	A custom provider, openai-completions	You front another model API

Three ways to set it up

From Ollama. Ollama’s own docs give one command: ollama launch openclaw. It offers to install OpenClaw through npm if it is missing, shows a security notice about tool access, lets you pick a local or cloud model, configures the provider and starts the gateway. OpenClaw was previously called Clawdbot, and Ollama says ollama launch clawdbot still works. If you write the config by hand, OpenClaw’s docs are emphatic about one setting: point it at http://host:11434, because the /v1 address “breaks tool calling” and models start printing tool-call JSON as text.

From OpenClaw’s onboarding. Install the llama.cpp plugin, run openclaw onboard and choose Managed local server. The docs say setup shows the model, download size and backend before downloading, then verifies a real tool call before switching the default model.

From LM Studio or any OpenAI-style server. Add a provider to ~/.openclaw/openclaw.json with its baseUrl, the model ID, its context window and an api type. OpenClaw’s docs recommend openai-responses where the server supports it, as LM Studio does, and openai-completions otherwise. A placeholder API key such as ollama-local is accepted for servers on your own machine or LAN. If you need a first model to try, our starter guide covers installing Ollama and LM Studio.

You do not have to go fully local. The docs show a hybrid setup in which a local model is the primary and a hosted model is a fallback, or the other way round; OpenClaw switches per turn when a provider errors.

What a local model needs

Context. Ollama recommends at least 64k tokens of context for OpenClaw on local models, and OpenClaw’s managed llama.cpp recipes use 64K too. That is far above the window Ollama starts with on cards under 24GiB, which is the same wall local coding tools hit.

Memory. OpenClaw does not give one figure. Its Local models page says memory depends on the weights, context size, runtime and whatever else the machine is doing. The smallest managed llama.cpp recipe has an 8 GiB host-memory floor, and the docs add that these floors “do not guarantee fit or speed”. For sizing a model to your own graphics card, see how much VRAM a model needs.

Working tool calls. Most of the docs’ local-model troubleshooting is about these. “A model that loads or answers a short prompt may still fail an agent turn,” they warn. When tool calls come back as raw JSON, XML or an empty list, the advice is to fix the server’s chat template and parser first. OpenClaw can force tool use per model with tool_choice: "required", and as a last resort switch tools off for that model, so the agent runs without tool calls. For Ollama, LM Studio and managed servers it turns on Tool Search automatically, which defers tool definitions to save context.

Security is your job, more than usual

OpenClaw’s Local models page is blunt: “Local models do not provide hosted providers’ safety filters.” Its prompt-injection page goes further, saying smaller models are “more susceptible to tool misuse and instruction hijacking”, and that for agents that run tools or read untrusted content, the risk with older or smaller models “is often too high”.

If you run a small local model anyway, the docs’ list is specific: read-only tools, sandboxing for every session, minimal filesystem access, strict allowlists, and web search, web fetch and the browser switched off unless inputs are tightly controlled. Sandboxing is opt-in, so it is off until you enable it. For a chat-only assistant with trusted input and no tools, the docs say smaller models are usually fine.

The defaults help. On a normal install the gateway listens only on your own machine, and most chat channels answer an unknown sender with a pairing code rather than acting on the message. openclaw security audit reports where your setup has drifted from that.

Questions people ask

Can OpenClaw run with a local LLM?

Yes. OpenClaw’s docs support local models through Ollama, LM Studio, an OpenClaw-managed llama.cpp server, and any server with an OpenAI-style /v1/chat/completions endpoint, such as vLLM, SGLang or MLX.

How do I use OpenClaw with Ollama?

Run ollama launch openclaw. Ollama’s docs say it installs OpenClaw if needed, shows a security notice, lets you pick a model, configures the provider and starts the gateway. If you configure it by hand, use the native address http://host:11434, not /v1.

What context window does OpenClaw need?

Ollama recommends at least 64k tokens for local models with OpenClaw. OpenClaw’s own managed llama.cpp recipes also use a 64K context, and its docs say to leave room for the full prompt, tools, history and output.

Why does my local model print tool-call JSON instead of running tools?

OpenClaw’s docs give two causes: pointing it at Ollama’s /v1 address, which breaks tool calling, or a server whose chat template and parser do not support tool calls. Fix the template first; forcing tool_choice is a per-model override.

Does OpenClaw work with LM Studio?

Yes. Start LM Studio’s local server (default http://127.0.0.1:1234), add it as a provider in ~/.openclaw/openclaw.json, and use the Responses API, which OpenClaw’s docs recommend so reasoning stays separate from the final reply.

Is it safe to run OpenClaw on a small local model?

OpenClaw’s docs advise against small models for agents that run tools or read untrusted content. If you must, they say to use read-only tools, sandboxing, minimal file access, strict allowlists, and to switch off web search, web fetch and the browser.