Skip to content

OpenClaw Additional Providers: DeepSeek, Groq, Ollama, OpenRouter, Bedrock...

Mar 28, 2026 1 min
Home › ai
TL;DR Beyond the big three (Anthropic/OpenAI/Google), OpenClaw supports 30+ providers — from DeepSeek to local Ollama and everything in between.
#openclaw#deepseek#groq#ollama#openrouter#vllm#bedrock#sglang#mistral
Table of Contents
  1. Unified Configuration Format
  2. Cost-Effective: DeepSeek
  3. Ultra-Fast Inference: Groq
  4. Local Models: Ollama
  5. Local Models: vLLM
  6. Local Models: SGLang
  7. Unified Gateway: OpenRouter
  8. Enterprise Cloud: Amazon Bedrock
  9. Other Providers at a Glance
  10. Using Proxy Gateways
  11. Speech Transcription: Deepgram
  12. Community Tool: Claude Max API Proxy
  13. The Big Picture
  14. References

🌏 中文版

The previous model article covered Anthropic, OpenAI, and Google — the big three providers. But OpenClaw supports 35+ providers. This article covers the rest — from cost-effective DeepSeek, to ultra-fast Groq inference, to locally deployed Ollama and vLLM.

Unified Configuration Format

All providers follow the same pattern:

{
  env: { PROVIDER_API_KEY: "your-key" },
  agents: {
    defaults: {
      model: { primary: "provider/model-name" }
    }
  }
}

Most providers can be configured interactively with openclaw onboard.

Cost-Effective: DeepSeek

A Chinese AI company with an OpenAI-compatible API at extremely low prices.

openclaw onboard --auth-choice deepseek-api-key
ModelUse CaseContext
deepseek-chat (V3.2)General conversation128K
deepseek-reasoner (V3.2)Reasoning / chain-of-thought128K

Set the environment variable DEEPSEEK_API_KEY. If Gateway runs as a daemon, make sure the key is in ~/.openclaw/.env.

Ultra-Fast Inference: Groq

Groq uses its proprietary LPU hardware to run open-source models with extremely fast inference speeds.

{
  env: { GROQ_API_KEY: "gsk_..." },
  agents: {
    defaults: {
      model: { primary: "groq/llama-3.3-70b-versatile" }
    }
  }
}

Commonly used models:

ModelHighlights
Llama 3.3 70B VersatileBroad capabilities, long context
Llama 3.1 8B InstantSpeed-oriented
Gemma 2 9BLightweight
Mixtral 8x7BMoE architecture, complex reasoning

Bonus feature: Groq’s Whisper can do fast speech transcription — just configure it as the media-understanding provider.

Local Models: Ollama

Run open-source models on your own machine at zero cost.

# Install Ollama
# Pull a model
ollama pull glm-4.7-flash

# OpenClaw configuration
openclaw onboard  # Select Ollama

OpenClaw automatically discovers local Ollama models. It also supports cloud models (kimi-k2.5:cloud, minimax-m2.5:cloud, etc.).

Important warning: Do not use the /v1 OpenAI-compatible URL. This breaks tool calling — the model will output tool JSON as plain text. Use the native Ollama API URL: http://host:11434 (without /v1).

Minimal setup: Set OLLAMA_API_KEY=ollama-local, and OpenClaw handles auto-discovery.

Local Models: vLLM

Serves open-source and custom models via an OpenAI-compatible HTTP API.

# Start vLLM server
# Configure
export VLLM_API_KEY="vllm-local"
{
  agents: {
    defaults: {
      model: { primary: "vllm/your-model-id" }
    }
  }
}

Like Ollama, vLLM supports auto-discovery — if VLLM_API_KEY is set but no provider config is specified, OpenClaw will query GET http://127.0.0.1:8000/v1/models.

Manual configuration allows specifying parameters like contextWindow, maxTokens, etc.

Local Models: SGLang

Another local model runtime. Documentation is in docs/providers/sglang.md, and the configuration approach is similar to vLLM.

Unified Gateway: OpenRouter

One API key to access models from multiple providers. Model format: openrouter/<provider>/<model>.

openclaw onboard --auth-choice apiKey --token-provider openrouter --token "$OPENROUTER_API_KEY"
{
  agents: {
    defaults: {
      model: { primary: "openrouter/anthropic/claude-sonnet-4-6" }
    }
  }
}

Ideal for those who want a single bill for accessing models from multiple providers. OpenRouter automatically injects cache control for Anthropic models.

Enterprise Cloud: Amazon Bedrock

No API key needed — uses the AWS SDK default credential chain.

{
  models: {
    providers: {
      "amazon-bedrock": {
        api: "bedrock-converse-stream"
        // No apiKey needed, uses AWS credentials
      }
    }
  }
}

Authentication priority: AWS_BEARER_TOKEN_BEDROCK → standard AWS credentials → profiles → SDK chain.

Supports auto-discovery — if AWS credentials are available, it automatically lists available Bedrock models (via bedrock:ListFoundationModels, cached for 1 hour).

When using instance roles on EC2, set AWS_PROFILE=default to tell OpenClaw that credentials are available.

Required IAM permissions: bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, bedrock:ListFoundationModels.

Other Providers at a Glance

Each provider has its own dedicated documentation. Here are the highlights:

ProviderProvider IDAuth MethodHighlights
MistralmistralAPI KeyEuropean company, strong multilingual support
xAIxaiAPI KeyGrok models
NVIDIAnvidiaAPI KeyNIM inference service
Hugging FacehuggingfaceAPI KeyInference API
Together AItogetherAPI KeyMulti-model inference platform
Qwen / Alibaba Cloudqwen_modelstudioAPI KeyQwen series
GLM (Zhipu AI)glmAPI KeyGLM series
MiniMaxminimaxAPI KeyChinese AI
Moonshot (Kimi)moonshotAPI KeyKimi series, includes Kimi Coding
Qianfan (Baidu)qianfanAPI KeyERNIE Bot
Volcengine (Doubao)volcengineAPI KeyByteDance
XiaomixiaomiAPI KeyXiaomi AI
VeniceveniceAPI KeyPrivacy-focused
GitHub Copilotgithub-copilotOAuthUses Copilot subscription
LiteLLMlitellmCustomUnified proxy gateway
Vercel AI Gateway—CustomVercel proxy
Cloudflare AI Gateway—CustomCloudflare proxy

Using Proxy Gateways

LiteLLM, Vercel AI Gateway, and Cloudflare AI Gateway are not model providers — they are proxy layers. You place them between OpenClaw and the actual provider for unified billing, rate limiting, or routing.

Speech Transcription: Deepgram

Not a language model, but a speech transcription service. Once configured as the transcription provider, voice messages are automatically converted to text.

Community Tool: Claude Max API Proxy

A community-maintained proxy that uses Claude subscription credentials to access the API. Make sure to verify compliance with Anthropic’s Terms of Service before using it.

The Big Picture

OpenClaw’s provider ecosystem breaks down into four tiers:

  1. Top-tier commercial (Anthropic / OpenAI / Google) — Strongest capabilities, most expensive
  2. Cost-effective (DeepSeek / Groq / Mistral) — Cheap or with free tiers, solid capabilities
  3. Local deployment (Ollama / vLLM / SGLang) — Zero cost, but requires your own hardware
  4. Proxy gateways (OpenRouter / LiteLLM) — One key to access multiple providers

You can mix and match across tiers in model.fallbacks — for example, use Claude as the primary, fall back to DeepSeek, then fall back to local Ollama.

References

This article is compiled from the following OpenClaw source documents:

Related · #openclaw

2026 LLM Inference Provider Free Tiers & Pricing: 40+ Services Ranked by Tier 1 min 2026 Q1 Open-Source LLM Landscape: From Frontier Models to On-Device, a Complete Survey 1 min OpenClaw Automation (Part 2): Standing Orders — Permanent Directives 1 min