Guide

OpenClaw Supported Models: Providers, Local Runtimes and How to Choose

2026-11-06

OpenClaw Supported Models: Providers, Local Runtimes and How to Choose

OpenClaw is not tied to one model. It connects to hosted providers such as Anthropic, OpenAI, Google Gemini, Groq, Mistral, DeepSeek, xAI, OpenRouter, Venice and Amazon Bedrock. It also connects to local runtimes such as Ollama, LM Studio, llama.cpp and vLLM, and to any OpenAI-compatible endpoint you define yourself. The official provider directory is the list of record. This guide explains how that list works in practice and how to pick a model for the job you want done.

TL;DR

  • OpenClaw uses provider plugins with their own model catalogs, plus custom providers for any compatible endpoint, so the "model list" is really a list of providers.
  • Hosted providers are the simplest path. Local runtimes (Ollama, LM Studio, llama.cpp, vLLM and others) work too, but they need enough RAM or VRAM and give you no hosted safety filters.
  • Set one strong primary model and add cheaper fallbacks. Avoid older, weaker tiers for agents that use tools.
  • If you want privacy, look at providers with no-logging tiers, and run your agent on an isolated, encrypted server.
  • On ClawBase, a managed dedicated server runs OpenClaw for you with access to over 50 AI models, so you choose models, not infrastructure.

Table of Contents

How OpenClaw handles models

It helps to drop the idea of a fixed list. OpenClaw connects through provider plugins that own their model catalogs, or through custom providers you configure under models.providers. A new model from a provider can show up without OpenClaw itself changing, because the catalog belongs to the provider plugin.

Setup is two steps, according to the model provider quickstart: authenticate with the provider (usually through openclaw onboard), then set the default model in config. Models are always written as provider/model. The quickstart example uses anthropic/claude-opus-4-6.

That provider/model format is the thing to learn first. Once you can read it, every entry in the directory makes sense, and switching models is a one-line change.

Hosted providers OpenClaw supports

The provider directory lists, among others, Amazon Bedrock, Anthropic, DeepSeek, Google Gemini, Groq, Mistral, Ollama, OpenAI, OpenRouter, Venice and xAI (OpenClaw provider directory). Two entries show how the format works in practice.

ProviderProvider IDAuth variableDefault / example models
OpenAI`openai``OPENAI_API_KEY`Fresh setup default `openai/gpt-6-astra`; examples include gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna and gpt-5.5
Google Gemini`google``GEMINI_API_KEY`Default `google/gemini-4-argon`; bundled models include `google/gemini-3.8-flash` and `google/gemini-3.5-flash-lite`

Both rows come from the official provider plugins page. Model names change fast, so treat any table, including this one, as a snapshot and check that page for the current catalog.

Gateways and aggregators

OpenRouter sits in the directory as a single endpoint in front of many models. That is convenient when you want to try several without opening several accounts. Amazon Bedrock fills the same role for teams already inside AWS. In both cases the provider ID changes, but the config pattern stays the same.

Privacy-oriented providers

Venice is worth a look if prompts are sensitive. It offers Private models, whose prompts and responses are never stored or logged, and Anonymized models, which are proxied through Venice with metadata stripped. Anonymized models are not fully private because the underlying provider still processes the request, so use Private models when you need full privacy. The built-in catalog includes 10 private and 6 anonymized models. The default is venice/zai-org-glm-4.7, and the strongest anonymized option is venice/claude-opus-5 (Venice AI in OpenClaw docs).

Local and self-hosted models

The directory also covers local options: Ollama (cloud and local models), Ollama Cloud, oMLX for Apple Silicon, llama.cpp, llmman, LM Studio, SGLang, vLLM and ds4 for local DeepSeek V4 (provider directory). The local models guide adds MLX, LiteLLM and custom OpenAI-compatible servers to the supported list.

A few details save time:

  • Ollama uses its native API. OpenClaw talks to /api/chat, not the OpenAI-compatible /v1 endpoint. It supports Cloud + Local, Local only, and Ollama Cloud through the ollama-cloud provider (Ollama in OpenClaw docs).
  • Memory has a floor. Managed llama.cpp recipes need at least 8 GiB of host memory, and hardware-aware setup checks RAM, GPU memory and disk before it picks a recipe (local models guide).
  • Custom servers need two things: a base URL and a model ID.

How much hardware do you need?

A community hardware guide gives rough tiers: 8 GB of RAM or VRAM handles 7-8B models for daily assistant use, 16 GB handles 14B models for tool-calling agents, 24 GB of unified memory on an M-series Mac handles quantized 32B models for serious coding, and 32 GB or more handles 32B full or 70B quantized (HowOpenClaw hardware guide). Treat those as starting points, not guarantees.

Hardware memory tiers and model sizes

One safety point belongs here. The OpenClaw docs note that local models do not carry hosted providers' safety filters, so keep tool permissions and prompt-injection defenses appropriate for the model and the task (local models guide). If your agent can read files or send messages, scope its permissions tightly before you point it at a small local model.

Custom providers: any OpenAI-compatible endpoint

When a service is not in the directory, you add it yourself. A custom endpoint takes a baseUrl, an apiKey and an api type. If you leave out api and set a baseUrl, the provider defaults to openai-completions (custom providers docs).

The configuration reference lists the supported API types: openai-completions, openai-responses, openai-chatgpt-responses, anthropic-messages, google-generative-ai, google-vertex, github-copilot, bedrock-converse-stream, ollama, pi-messages and azure-openai-responses (custom providers configuration).

Two behaviors are useful to know when you debug a custom setup:

  1. On non-native OpenAI-compatible routes, OpenClaw forces compat.supportsDeveloperRole: false to avoid provider 400 errors.
  2. Proxy-style routes skip native OpenAI-only request shaping, such as service_tier, Responses store, Completions store, prompt-cache hints and hidden attribution headers (custom providers docs).

The baseUrl is also a trust decision. OpenClaw allows that exact scheme, host and port through its guarded fetch path and does not trust other private origins as a side effect (custom providers configuration).

Using OpenClaw as an OpenAI-style endpoint

The connection also runs the other way. The Gateway can expose an OpenAI-compatible Chat Completions surface at /v1/chat/completions, /v1/models and /v1/embeddings. It is disabled by default, and each request runs as a normal Gateway agent run (OpenAI chat completions docs). That lets existing tools that speak the OpenAI format talk to your agent, with its memory and skills behind it.

If you are building a product on top of that, say an internal tool or a customer-facing assistant, teams often bring in specialists in AI product design and development to shape the interface around the agent.

Choosing a model for the job

OpenClaw's own model policy is short and practical: set the primary model to the strongest latest-generation model you have access to, use fallbacks for cost- or latency-sensitive work, and avoid older or weaker tiers for agents that have tools enabled or read untrusted input (Models CLI docs). Everything below builds on that.

A simple three-slot setup

  1. Primary: a top-tier model for planning, tool calls and anything touching files or accounts.
  2. Fallback: a faster, cheaper model for routine replies, summaries and reminders.
  3. Local or private option: a self-hosted or no-logging model for sensitive material.

This maps onto what people use agents for. Everyday assistance (inbox triage, scheduling, quick answers) rarely needs the top tier. Coding and multi-step automation do, because a weak model that mis-calls a tool costs you more in cleanup than the tokens you saved. For ideas on what to automate, see our OpenClaw use cases.

What about free and low-cost models?

Search results for this topic are full of "best free model" lists, and free options do exist. Two published examples:

  • Groq's free tier has included Llama 3.1 8B, Llama 3.3 70B, Llama 4 Scout and GPT-OSS 120B, with published limits of 30 requests per minute, 6,000 tokens per minute and 14,400 requests per day (Findstack).
  • OpenRouter's free tier, as one writer counted it in May 2026, offered 29 :free models at 20 requests per minute and 200 requests per day (Thomas Cherickal).

Free tiers change often, so confirm the current limits with the provider. They suit experiments and fallback duty. An always-on agent that runs scheduled jobs will hit a per-day cap quickly, so keep a paid primary for anything you depend on.

On cost overall, OpenClaw itself is MIT-licensed and free. The running cost is mostly model API usage, which one ranking site estimates at roughly $6 to $200 or more per month depending on use, plus any server you host it on (Synthszr). Heavy tool use and long context drive that figure more than the sticker price per token.

A quick decision guide

If you want...Look at
The simplest startA single hosted provider (Anthropic, OpenAI or Google) with its default model
Many models, one keyAn aggregator such as OpenRouter
Lowest cost for light useA fast tier from a hosted provider, or a free tier as a fallback
Offline or data-on-your-hardware useOllama, LM Studio, llama.cpp or vLLM, sized to your RAM
Strong privacy on hosted modelsVenice Private models
An endpoint nobody listsA custom provider with `baseUrl` and `apiKey`

Alternatives to document-chat tools

People who search for alternatives to local document-chat projects usually want one of three things: a chat interface over their files, a self-hosted chat front end for many models, or an agent that acts rather than only answers. OpenClaw sits in the third group. It runs tasks, manages files, remembers context between sessions and connects to messaging apps, while still letting you point it at local models if you want your documents to stay on your own hardware. If your goal is only to query a folder of PDFs, a simple retrieval tool may be enough. If you want an assistant that also does things on your behalf, an agent like OpenClaw fits better.

Perspective: why we think model choice should not be a sysadmin task

> We built ClawBase because we kept seeing the same pattern. People are excited about OpenClaw, then lose a weekend to servers, keys, config files and updates before the assistant answers a single question.

>

> Our view is that the interesting decisions are about models and memory, not infrastructure. Which model should plan, which should handle the cheap routine work, how much context should persist between sessions: those are the choices that change what your assistant can do. Patching a server is not.

>

> That is why ClawBase deploys OpenClaw in one click on a dedicated, encrypted server, with persistent memory, daily encrypted backups and automated updates, and gives you access to over 50 AI models, including multi-model routing. You still decide which models do the thinking. We take care of where they run.

>

> Our advice stays close to OpenClaw's own: put a strong model on anything with tools, use cheaper fallbacks for the rest, and keep sensitive work on a private path. A private, always-on server makes that easy to hold to, because your agent and its memory are not shared with anyone else.

>

> *Iosif Peterfi*

Run OpenClaw with the models you choose

If you would rather pick models than manage servers, ClawBase gives you a managed OpenClaw on its own dedicated, encrypted cloud server. It stays on around the clock (99.9% uptime), connects to Telegram, Discord, Slack and WhatsApp, and needs no sysadmin skills.

Plans are LITE at $16 per month or $199 per year, PRO at $33 per month or $399 per year, and MAX at $66 per month or $799 per year. The entry plan includes a 7-day free trial, and you can pay by credit card or cryptocurrency. See the ClawBase pricing page for what each plan includes.

Clawbase

If you handle regulated data, read how a private managed OpenClaw server fits U.S. teams. If a setup error is the thing slowing you down, our open-source AI setup errors guide covers the common ones.

FAQ

Which AI models does OpenClaw support?

OpenClaw supports models from many providers, including Anthropic, OpenAI, Google Gemini, Groq, Mistral, DeepSeek, xAI, OpenRouter, Venice and Amazon Bedrock. It also supports local runtimes such as Ollama, LM Studio, llama.cpp and vLLM. Anything else that speaks a compatible API can be added as a custom provider.

Where is the official OpenClaw model list?

There is no single fixed list, because each provider plugin owns its catalog. The provider directory shows the supported providers, and the official provider plugins page shows each provider's default and example models.

Can OpenClaw run models locally?

Yes. It works with Ollama, LM Studio, llama.cpp, vLLM, SGLang, MLX, LiteLLM and custom OpenAI-compatible servers. Managed llama.cpp recipes need at least 8 GiB of host memory. Local models lack hosted safety filters, so limit tool permissions accordingly.

How do I add a model that is not listed?

Define a custom provider under models.providers with a baseUrl, an apiKey and, if needed, an api type. With only a baseUrl, it defaults to openai-completions. Then set the model as provider/model in your config.

Can I use more than one model at a time?

Yes. Set a primary model and add fallbacks for cheaper or faster tasks, as OpenClaw's model policy suggests. On ClawBase you can route across the models available on your plan.

Sources

Recommended