Groq
FreeRecommended firstFast daily chat and first-time setup
- You need
- A free API key, no credit card
- Models
- Every chat model Groq currently serves.
- GPT-OSS 120B (default)GPT-OSS 20BGroq CompoundGroq Compound Mini
AI CoreChoose the AI that
Choose the AI that
thinks for Jarvis
AI Core is where you connect Jarvis to an AI model: a cloud provider with your own key, a Claude plan you already have, or a model running entirely on your PC. You also choose how private every request should be.
8
AI providers
5 cloud, 1 subscription, 2 on your PC
Live
Model lists
new models appear without an app update
4
Built-in local models
for the llama.cpp engine
4
Privacy modes
from Standard to This PC only
Providers
8 providers, one screen
Each provider is a card in AI Core: click one to switch. Jarvis asks each provider which models it serves right now, so the list is always current and new models appear without an app update. The models below are examples.
Claude
PaidReasoning, writing, planning and careful answers
- You need
- An Anthropic API key
- Models
- Every model your Anthropic account can use.
- Claude Opus 4.8 (default)Claude Opus 5Claude Sonnet 5Claude Sonnet 4.6Claude Haiku 4.5
Claude Code
SubscriptionTop-tier reasoning through a Claude plan you already pay for
- You need
- The Claude Code CLI signed in on this PC; no API key
- Models
- Detected automatically when the CLI is installed.
- OpusSonnetHaikuSonnet 5Opus 4.8
OpenAI
PaidGeneral-purpose chat and broad compatibility
- You need
- An OpenAI API key
- Models
- Every chat model your OpenAI account can use.
- GPT-4o (default)GPT-5GPT-5 miniGPT-4.1o4-mini
Gemini
Free tierA free-tier option and long documents
- You need
- A Google AI Studio key (free tier available)
- Models
- Every Gemini model that can chat.
- Gemini 3.6 Flash (default)Gemini 3 Flash (preview)Gemini Flash Lite
OpenRouter
PaidTrying many model families from one account
- You need
- An OpenRouter API key (some models are free)
- Models
- Up to 100 models from OpenRouter’s catalogue in the picker.
- Free models such as Nemotron 3 Super, Qwen3 Coder and DeepSeek ChatGPT-4oClaude 3.5 Haiku
Ollama
LocalPrivate setups, or free cloud models through Ollama
- You need
- Ollama installed; a free ollama.com account for cloud models
- Models
- Any model you have pulled, plus suggestions that fit your PC.
- Cloud: GPT-OSS 20B / 120B, Gemma 4, MiniMax M2.5, Nemotron 3 SuperLocal: Llama 3.2 3B, Qwen 2.5 7B, Llama 3.1 8B, Mistral Nemo 12B
llama.cpp
On-deviceOffline, fully private answers with nothing extra to install
- You need
- One-click install inside Jarvis (about 50 MB), then a model
- Models
- Four tested models, offered only if your PC can run them.
- Qwen 3.5 9B (default)Qwen 3.5 2BGemma 4 12BMuse Glimmer 30B
Local models
The built-in llama.cpp engine
Run an AI model entirely on your PC, with nothing extra to install and nothing sent to the cloud. Jarvis offers four tested models and only shows the ones your hardware can run.
- Install from AI Core: the engine is about 50 MB, and nothing downloads until you click.
- Jarvis checks your graphics card, memory and disk first, and only offers models your PC can run.
- Each model’s size, publisher and licence (Apache 2.0) are shown before it downloads.
- Uses NVIDIA (CUDA) or AMD and Intel graphics (Vulkan), and falls back to the CPU.
- Needs a 64-bit CPU with AVX2 and at least 8 GB of RAM.
| Model | Download | Graphics card | Context | Can do |
|---|---|---|---|---|
| Qwen 3.5 2B Small and quick, but too small to use tools reliably, so it is labelled chat only. | 1.3 GB | None needed (CPU or integrated graphics) | 32K | Chat only |
| Qwen 3.5 9B The default. Fits the widest range of PCs that can use tools. | 6.0 GB | 8 GB free video memory | 32K | Chat + tools |
| Gemma 4 12B | 7.4 GB | 12 GB free video memory | 32K | Chat + tools |
| Muse Glimmer 30B The most capable. Comes with an extra usage policy, shown before you download. | 15.9 GB | 20 GB free video memory | 128K | Chat + tools |
Prefer Ollama? It works too: Jarvis suggests local models that fit your PC, labels which can use tools, and also offers Ollama's free cloud models after you sign in at ollama.com.
Manual or Auto
You pick the model, or Jarvis does
Every setting saves the moment you change it.
Manual
DefaultJarvis always uses the model you picked.
Auto
Jarvis picks a model for you, from the providers you tick. Choose what it optimises for: Balanced, Faster or Higher quality. Ollama and Claude Code stay off unless you tick them, and Auto only picks free models until you allow paid ones.
Privacy
Decide what leaves your PC
Four modes, changed in one click and applied straight away, with no restart. Jarvis asks you to confirm before switching to a stricter or more permissive mode, and tells you exactly what changes.
Standard
RelaxedDefaultEverything works. Your messages go to the AI model you picked, exactly as you wrote them.
Redact secrets
GuardedSame models and features, but anything that looks like an API key, password or access token is stripped out before your message leaves your PC.
This PC only
StrictNo prompt reaches a cloud AI model. Every cloud provider is switched off, even with a key saved, and only local models (Ollama or llama.cpp) answer. Online tools like web search still work.
Full access
Everything in Standard, and Jarvis stops asking before it acts. Every action is still recorded.
Emergency Stop, anything you set to "never allow" and the built-in blocks on dangerous commands still apply.
Advanced
Fine-tune it, if you want to
Most people never need to open Advanced. It is there when you want more control.
Fallback models
If your model fails, Jarvis retries with these, in the order you set. In Manual mode it only uses the list if you turn that on.
Per-task models
Point a specific job at a different model; everything else keeps your main model.
- • Background tasks: reminders, reflections and scheduled jobs
- • Browser: when Jarvis browses the web for you
- • Code review: reviewing code changes and pull requests
- • Classifier: small internal decisions, like sorting messages
Privacy for email and calendar
Optional rules that never turn on by themselves.
- • Prefer a private model: try a chosen model first for email and calendar content
- • Require on-device: email and calendar content never reaches a cloud model
Ollama settings
Cloud-only mode is recommended. Turn on local models and Jarvis shows how well your PC can run them, warns you first, and only offers models that fit.
AI Core · AdvancedScreenshot coming soon
Get started
Three ways to set it up
Free and fast
Best for most people starting out
- 1Open AI Core and choose Groq.
- 2Get a free key at console.groq.com and paste it in.
- 3Jarvis checks the key works, then saves it. Done.
Best quality
For stronger reasoning and writing
- 1Choose Claude and paste an Anthropic API key, or
- 2Install and sign in to the Claude Code CLI to use your Claude plan with no key.
- 3Pick a model from the list.
Fully private
Your prompts stay on your PC
- 1Choose llama.cpp and click Install engine.
- 2Download the model Jarvis recommends for your PC.
- 3Set privacy to This PC only.
FAQ
Common questions
+How many models can I use?
It depends on your providers. Jarvis asks each provider which models it serves right now, so the list is always current: every model your Claude, OpenAI, Gemini or Groq account can use, up to 100 from OpenRouter, any model you have in Ollama, and four tested models for the built-in llama.cpp engine.
+Do I need to pay for an AI model?
No. Groq has a free key, Gemini has a free tier, OpenRouter and Ollama offer free models, and local models through llama.cpp or Ollama cost nothing to run. Paid providers bill you directly; Jarvis adds no markup.
+Can I use more than one provider?
Yes. Connect as many as you like and click a card to switch. Auto mode can choose between them for you, and fallback models step in if one fails.
+Where are my API keys stored?
On your PC, encrypted with Windows data protection. They are only sent to the provider they belong to.
+What happens if a provider goes down?
Jarvis retries with your fallback models, in order. Set them under Advanced → Fallback models.