Skip to main content
At a Glance

Two Ways to Run OpenClaw

Choose based on your budget, privacy preferences, and how much you want to manage yourself.

🖥️
$0
Local Model
Run Ollama or LM Studio on your machine. No API key, no monthly bill. Slower and less capable, but completely private and free.
🔑
$5–30/mo
Self-Hosted + API
Run OpenClaw yourself, bring your own LLM API key. You control the hardware. Pay only for what you use.
⚠️

The most common beginner mistake: Assuming OpenClaw has a subscription. It doesn't. The "cost" beginners hit is when they try to use Claude or GPT-4 and realise they need an API key. You can avoid this entirely by starting with a free local model, then upgrading when you want better quality.

LLM APIs

Which LLM Should You Use?

OpenClaw works with any OpenAI-compatible API. Here's how the major options compare on price, quality, and speed. All prices as of early 2026 — check provider sites for current rates.

Model Provider Input cost Output cost Typical monthly Best for Quality
claude-sonnet-4 Anthropic $3 / M tokens $15 / M tokens $10–35 Complex tasks, coding, reasoning ⭐⭐⭐⭐⭐
claude-haiku-4 Anthropic $0.80 / M tokens $4 / M tokens $3–12 Fast daily tasks, high volume ⭐⭐⭐⭐
gpt-4o OpenAI $2.50 / M tokens $10 / M tokens $8–30 General purpose, wide tool support ⭐⭐⭐⭐⭐
gpt-4o-mini OpenAI $0.15 / M tokens $0.60 / M tokens $1–5 Budget option, simple tasks ⭐⭐⭐
gemini-2.0-flash Google $0.10 / M tokens $0.40 / M tokens $1–4 Speed, large context, low cost ⭐⭐⭐⭐
gemini-1.5-pro Google $1.25 / M tokens $5.00 / M tokens $5–20 1M token context window ⭐⭐⭐⭐
Llama 3.3 70B Ollama (local) Free Free $0 Privacy, offline use, no API key ⭐⭐⭐
Mistral 7B Ollama (local) Free Free $0 Fast local inference on modest hardware ⭐⭐⭐
💡

Recommended starting point: claude-haiku-4 or gpt-4o-mini for most beginners. They deliver 90% of the quality at 10% of the cost. Upgrade to Sonnet or GPT-4o for complex reasoning tasks or coding projects.

What's a "token"?

A token is roughly ¾ of a word. "Hello, how are you?" is about 5 tokens. A typical conversation message is 100–500 tokens. A long document summary might be 5,000 tokens. For most OpenClaw users, a full day of active use generates around 50,000–200,000 tokens — which costs $0.04–$3.00 depending on your model choice.

# Set your LLM in ~/.openclaw/openclaw.json
{
  "llm": {
    "provider": "anthropic",        // or "openai", "google", "ollama"
    "model":    "claude-haiku-4",    // model name from provider docs
    "apiKey":   "sk-ant-..."         // your API key (keep this secret)
  }
}
Cost Estimator

What Will You Actually Pay?

Drag the sliders to match your usage. Costs are estimates based on average conversation length and API pricing.

Monthly Cost Estimator

Adjust sliders to match your typical usage. Estimates assume ~400 tokens/message average.

Messages per day
How many times you send a message to your agent
30
Automations per day
HEARTBEAT tasks and background runs (typically 500–2000 tokens each)
5
Model
Higher quality = higher cost
Estimated monthly cost
Based on 400 tokens/message avg, 30-day month
$3.84
🎉

Most users land between $3–15/month. Even heavy users running dozens of automations daily rarely exceed $30 with a mid-tier model. The only way to rack up large bills is running many long document processing tasks or extremely high-volume automations without caching.

Run It Free

$0/Month with Local Models

You can run OpenClaw completely free by using a local language model on your own hardware. No API key, no monthly bill, no data leaving your machine.

🦙

Ollama

The easiest way to run local models. One-line install, supports Llama, Mistral, Qwen, Gemma, and 50+ models. Works on macOS, Windows, and Linux.

Free · Offline · Private
🏠

LM Studio

Desktop GUI for running local models. Great for beginners who want a visual interface to download and manage models. Exposes a local API endpoint compatible with OpenClaw.

Free · GUI · macOS/Windows
⚡

Jan.ai

Open-source alternative to ChatGPT you run locally. Includes a built-in API server. Particularly good performance on Apple Silicon Macs via Metal acceleration.

Free · Fast · Apple Silicon
🌐

Groq (free tier)

Cloud-hosted inference with a generous free tier (14,400 requests/day). Extremely fast (400+ tokens/sec). Not fully private but costs nothing for typical use.

Free tier · Very fast · API key needed

Setting up Ollama with OpenClaw

# Step 1: Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Step 2: Pull a model (Llama 3.3 recommended for most tasks)
ollama pull llama3.3

# Step 3: Configure OpenClaw to use it (~/.openclaw/openclaw.json)
{
  "llm": {
    "provider": "ollama",
    "model":    "llama3.3",
    "baseUrl":  "http://localhost:11434"
  }
}
💻

Hardware requirements: For a usable local model, you need at least 8 GB RAM for a 7B model and 16 GB RAM for a 13B model. Apple Silicon Macs (M1/M2/M3/M4) run local models especially well thanks to unified memory. Local models are noticeably slower and less capable than GPT-4o or Claude Sonnet — but for simple daily tasks, automation scripts, and reminders, they're more than adequate.

Best free local models for OpenClaw tasks

Model Size RAM needed Best for Speed
Llama 3.3 70B 40 GB 48 GB+ Complex reasoning, coding ⭐⭐⭐
Llama 3.2 8B 5 GB 8 GB Daily tasks, reminders, email ⭐⭐⭐⭐⭐
Mistral 7B 4 GB 8 GB Fast responses, summarisation ⭐⭐⭐⭐⭐
Qwen2.5 14B 9 GB 16 GB Coding, data tasks, multilingual ⭐⭐⭐⭐
Gemma 3 12B 8 GB 16 GB Balanced quality and speed ⭐⭐⭐⭐
Save Money

5 Ways to Lower Your API Bill

Even small tweaks can cut your LLM costs by 50–80%.

  • ⚡

    Use a fast/cheap model for simple tasks

    Set your default model to Haiku or GPT-4o-mini. Only switch to Sonnet or GPT-4o when your AGENTS.md explicitly requests a power model for complex work. 80% of daily tasks don't need top-tier reasoning.

  • 📝

    Keep your MEMORY.md tight

    Every session loads your full MEMORY.md. A bloated 5,000-word memory file costs tokens on every single message. Aim for under 2,000 words and prune quarterly. Use /memory forget to remove outdated entries.

  • ⏱️

    Reduce heartbeat frequency for light tasks

    The default heartbeat runs every 30 minutes. If your scheduled tasks are light (weather checks, reminders), increase the interval to 60 or 120 minutes. Each heartbeat cycle uses 1,000–3,000 tokens.

  • 🗃️

    Enable prompt caching

    Anthropic and OpenAI both support prompt caching. With caching enabled, repeated context (like your AGENTS.md, USER.md, and system prompt) is cached and re-billed at 10% of the normal token cost.

  • 🔀

    Use local models for automations

    Configure a cheap or local model for HEARTBEAT automations and reserve your API model for interactive conversations. Add "heartbeatModel": "llama3.2" to your config to route background tasks to Ollama.

  • 📊

    Watch your usage dashboard

    Run openclaw stats to see your daily token usage broken down by conversation, automation, and model. Most API providers also have dashboards with spend limits you can set to avoid surprise bills.

# Route expensive tasks to Sonnet, background tasks to local Llama
{
  "llm": {
    "default":         "claude-haiku-4",
    "heartbeatModel":  "ollama/llama3.2",   // free for automations
    "powerModel":      "claude-sonnet-4"    // used when agent asks
  }
}
Side by Side

Local Model vs API: Which is Right for You?

Both options run on your own machine. The only difference is where the AI processing happens.

Feature Local Model (Ollama) Self-Hosted + LLM API
Monthly cost $0 $5–30 (API usage only)
Setup time ✓ ~5 minutes ✓ ~10 minutes
Response quality ⚠️ Good for daily tasks ✓ Excellent (Claude/GPT-4o)
Response speed ⚠️ Depends on your hardware ✓ Fast (cloud inference)
Data privacy ✓ Fully local, nothing leaves ⚠️ Sent to API provider
Works offline ✓ Yes ✗ Needs internet
Hardware needed ⚠️ 8–16 GB RAM recommended ✓ Any machine
Coding & complex tasks ⚠️ Capable but limited ✓ Excellent
Voice Mode ✓ Supported ✓ Supported
MCP integrations ✓ Unlimited ✓ Unlimited
💡

Our recommendation: Start with Ollama + Llama 3.2 to get familiar with OpenClaw at zero cost. Once you want better quality responses or are tackling more complex tasks, add a Claude Haiku or GPT-4o-mini API key. Most users make this switch within their first week.

FAQ

Pricing Questions Answered

OpenClaw the software is genuinely free and open-source (MIT licence). You can download it, modify it, and self-host it at no cost. The only cost you'll encounter is whatever LLM API you connect to it — which you'd be paying regardless of what AI tool you use. There's no hidden subscription, no feature gating, and no ads.
Yes, fully. Install Ollama on your computer, pull a free model like Llama 3.2, and point OpenClaw at your local Ollama endpoint. No API key, no credit card, no ongoing costs. The tradeoff is that local models are slower and less capable than Claude or GPT-4o — but for personal productivity tasks, reminders, and simple automation, they're very usable.
You can start with any of: an Anthropic API key (console.anthropic.com), an OpenAI API key (platform.openai.com), or a Google AI Studio key (aistudio.google.com). All three have free trial credits when you sign up. For the best experience, start with Claude Haiku or GPT-4o-mini — both are inexpensive and widely tested with OpenClaw.
This is very unlikely with normal usage, but you can protect yourself by setting a spend limit on your API account. Both Anthropic and OpenAI let you set a hard monthly cap. We also recommend starting with a $5 prepaid credit top-up rather than attaching a card directly. With Claude Haiku or GPT-4o-mini, $5 covers roughly 2–3 months of normal personal use.
Voice Mode uses two optional third-party services: OpenAI Whisper for speech-to-text (~$0.006/minute, or about $0.36/hour of voice) and ElevenLabs for text-to-speech (free tier: 10,000 characters/month; Starter plan: $5/mo for 30,000 characters). Many users find the free ElevenLabs tier more than enough. You can also use the free built-in native TTS on macOS/iOS as an alternative. See the Voice Mode pricing table for details.
The absolute cheapest path: install Ollama, pull the free Llama 3.2 8B model, and point OpenClaw at it. Total cost: $0. If you want better quality without spending much, sign up for an Anthropic API key, add $5 in prepaid credits, and use Claude Haiku — that $5 will typically last 2–3 months of personal use.

Start for Free Today

Install OpenClaw in 5 minutes, connect a free Ollama model, and see what your agent can do — before spending a cent.

Self-Host for Free → Compare to Alternatives