Every model. Several at once.
Manwe doesn't ship a model of its own. It runs the ones you choose: on your Mac, through the subscriptions you already have, or with any API key. Switch whenever you like; your Worlds, memory and agents stay put.
| Model | Runs | Needs | Status |
|---|---|---|---|
| Recommended local model | On your Mac | 8 GB+ memory | Recommended |
| Balanced local models | On your Mac | 16 GB+ memory | Experimental |
| Large local models | On your Mac | 32 GB+ memory | Experimental |
| Your MLX models | On your Mac | A linked model folder | Bring your own |
| Claude Sonnet, Opus | Claude Code | Your signed-in Claude Code | Subscription |
| GPT | Codex | Your signed-in Codex | Subscription |
| Anthropic, OpenAI, DashScope | API | Your API key | API |
| OpenRouter, Mistral, DeepSeek, Kimi, Groq, Together | API | Your API key | API |
| Ollama and loopback servers | On your Mac | An OpenAI-compatible server | Custom |
Requires an Apple silicon Mac with macOS 14 or later. Local models are a separate download. Third-party and custom models can produce formatting errors.
Choose how models run.
Pick per task, per advisor or per panel. A single question can go to a local model and two cloud models at once.
Local models
Supported MLX models run on Apple silicon, private and offline. The app recommends a responsive one to start with; follow its memory guidance for larger ones.
Subscription connections
Use your locally installed, signed-in Claude Code and Codex. Their account limits apply; they aren't included subscriptions.
API providers
Add your own providers and keys in Models & providers. Their capabilities, billing and limits apply. Keys live in macOS Keychain.
Know where the work happens.
Your workspace is stored on your Mac. The services you choose decide what needs to leave it, and only the context a task needs goes with it.
Model context
A cloud provider receives the relevant task context, which can include selected memories, documents and tool results. Different advisors can use different providers.
Local learning
Feedback signals and trained classifiers stay on your Mac. Learned preferences can become memory included in later model requests.
Online tools
A local model doesn't make research, connected services, maps or sharing offline. Secrets for APIs and connections are kept in Keychain.
Connect what the work needs.
Start with a model. Add folders, editors and apps when a task calls for them; each connection shows its setup and tools in Settings.
Files and editors
Grant a working folder or a file for the task. Use the included extension for VS Code or Cursor; Xcode needs Accessibility permission.
Apps and services
Connect supported services and local MCP apps. Calendar, Reminders, Contacts, Notes, Mail drafts, Maps, Photos and Music use their own macOS permissions.
Approvals
Approval modes control how actions proceed, and Manwe asks before anything you haven't allowed. macOS permissions stay separate.
Questions
Is there a Windows or Linux version?
Not today. Manwe is for Apple silicon Macs running macOS 14 or later.
What does a cloud model cost?
Your API provider or subscription connection decides its limits and charges. Separate services, such as image generation, can have their own costs.
Can I run GGUF models or my own runtime?
Yes, through an OpenAI-compatible server on your Mac, such as Ollama. Manwe connects to it; it doesn't launch the runtime for you.
Do I need Docker or an editor extension?
No. They're optional development features. Start with a model and add tools when a task needs them.
Bring your own models.
Version 0.9.93 · Apple silicon · macOS 14 or later
