When you create an assistant in HolaClaw, one of the first decisions is which AI model powers it. Each provider has its own personality, pricing, and trade-offs — and the right answer depends on what you're building.
The good news: you can change providers at any time, even for an assistant you've been using for months. Your memory, personality, and channels all stay intact.
Tip — the short version. If you're just starting out, pick Anthropic Claude. It's a great all-rounder, handles long conversations well, and works beautifully with HolaClaw's memory system. You can always switch later.
Quick comparison
Here's how the providers we support stack up at a glance. We'll go deeper into each below.
| Provider | Auth | Best for | Speed | Cost | Runs locally? |
|---|---|---|---|---|---|
| Claude | API key | Long conversations, nuance, writing | Fast | $$ | No |
| OpenAI | Subscription or API key | General use, tool calling, broad knowledge | Fast | $$ | No |
| Gemini | API key | Multimodal, very long context, low cost | Very fast | $ | No |
| Grok | API key | Snappy answers, distinct voice | Fast | $$ | No |
| DeepSeek | API key | Reasoning and chat at very low cost | Fast | ¢ | No |
| Local | — | Privacy, offline use, no API costs | Depends on Mac | Free | Yes |
Anthropic Claude
Claude is our default recommendation. It has a warm, conversational feel that pairs well with HolaClaw's memory system — it picks up on small details you mention and remembers them naturally across sessions.
Pick Claude if you want an assistant that holds long, thoughtful conversations. It's great for journaling companions, writing partners, and travel planners (like Luna).
Authentication: API key only. Anthropic doesn't offer a subscription that works outside of Claude.ai.
Available models
claude-sonnet-4.5— the sweet spot of speed and quality. Recommended default.claude-haiku-4.5— faster and cheaper, great for daily check-ins.claude-opus-4.1— the smartest model for complex reasoning, slower.
How to get a key. Sign in at console.anthropic.com, go to API Keys, and create a new key. Paste it into HolaClaw → Settings → AI Providers. We never see or store your key.
OpenAI
The most familiar option. GPT-4 and GPT-5 models are versatile, fast, and excellent at tool calling — useful if you're building an assistant that needs to look things up online or interact with other apps.
Pick OpenAI if you're already familiar with ChatGPT and want similar behavior, or if you need the broadest range of plugins and integrations.
Authentication: OpenAI is the only provider where you have two options:
- Sign in with your ChatGPT subscription and we'll route requests through it — no separate API billing.
- Or paste a platform API key from
platform.openai.comand pay per token. Useful if you want fine-grained model control or you don't have a paid ChatGPT plan.
Google Gemini
Gemini is the budget-friendly option with a huge context window — up to 2 million tokens. That's enough to feed it your entire library of past conversations and have it work with the lot.
Pick Gemini if you want low API costs, plan to send long PDFs or many images, or you're building an assistant that needs to "remember" months of context in one go.
Authentication: API key only. Get one from Google AI Studio (aistudio.google.com) and paste it into Settings.
Grok (xAI)
Grok has a distinct, looser voice and tends to answer with confidence on current events. The newer Grok models are fast and competitive with the mid-tier offerings from OpenAI and Anthropic.
Pick Grok if you want an assistant with a different personality from the default cloud models, or you're already paying for xAI access and want to route HolaClaw through it.
Authentication: API key only. Generate one in the xAI console and paste it into Settings.
DeepSeek
DeepSeek's models are remarkably capable for their price point — chat and reasoning models that cost a fraction of the equivalent OpenAI or Anthropic tier. Strong on math and code, and steadily improving on conversational tasks.
Pick DeepSeek if cost matters most and you're willing to trade a little polish for very cheap tokens, or you specifically want strong reasoning at low cost.
Authentication: API key only. Get one from platform.deepseek.com and paste it into Settings.
Local models
If privacy matters most, you can run a model entirely on your Mac. No data ever leaves your machine — no API keys, no cloud, nothing. HolaClaw ships with an embedded runtime, so there's nothing extra to install: open Settings → AI Providers → Local, pick a model, and you're done.
The trade-off is speed and quality. Local models are smaller than frontier cloud models, and they need an M-series Mac with at least 16GB of unified memory to feel responsive.
Heads-up — system requirements. Local models work best on M2 Pro / M3 / M4 Macs with 24GB+ of unified memory. On an 8GB Mac, expect slow responses and limited model choices.
Available local models
- Gemma 4 E2B — the smallest of the bunch. Snappy on any M-series Mac, good for quick check-ins.
- Gemma 4 E4B — a step up in quality with the same broad hardware support.
- Gemma 4 26B A4B MoE — a mixture-of-experts model that punches well above its activated parameter count. Best on 24GB+ machines.
- Gemma 4 31B — the strongest dense Gemma we ship. Needs 32GB+ unified memory to feel comfortable.
- GLM-4.6V Flash — fast and multimodal — handles images and screenshots alongside text.
- Ministral 3 14B Instruct — Mistral's instruction-tuned 14B. A solid all-rounder at moderate hardware cost.
Models are downloaded on demand from inside HolaClaw — no Terminal, no extra apps. You can swap models or remove their weights from disk at any time in Settings.
Bring your own key
For every cloud provider above, HolaClaw lets you bring your own credentials. API keys are stored in the macOS Keychain — encrypted, sandboxed, and never sent to our servers. We can't see your usage and don't bill you; everything goes directly to the provider.
Switching providers
You can swap providers for any assistant at any time. Open the assistant, click the provider badge in the header, and pick a new one. Your memory, personality, and conversation history come with you.
Switching mid-conversation is fine — the new model picks up where the previous one left off. The "voice" of the assistant may shift slightly, but the memory and facts remain.
Still not sure? Ask on Discord — the community is happy to share what they're using for what.