You pick the models. Every one.
Download a local model, connect one you host, or run cloud on your own key. Nothing is locked in.
A catalog, not a middleman
OpenShore starts you on three curated models, and the Marketplace, coming soon, will tell you in plain language what each model is good at. When you download, weights come straight from the source. OpenShore never rehosts weights or proxies your inference.
- Plain fit. Harbor Lite guides you on the phone, Harbor codes on the phone, and DeepBlue codes on your computer.
- Your machine's limit. DeepBlue is sized to what your computer can run, never the biggest on a guess.
- Licenses shown. All three curated models are Apache 2.0 open weights, and the app names the weights behind each one.
Models, in plain language
Curated for what each model is good at.
⌕Search models
All
Coding
Writing
Vision
This Mac runs up to ~32B
Qwen2.5 Coder 14BOn device
★★★★☆community
Coding · 9 GB
Details and license ›Get
Claude (frontier)On your key
Reasoning · cloud, your account
Only when you pick itConnect
Four ways to get a model
Mix and match. They all show up in Stack and place the same way.
- On device. Harbor Lite is built into the phone app and works offline. Harbor downloads when you want a coder on your phone, and DeepBlue comes through Ollama on your computer.
- Bring your own. Connect any OpenAI-compatible endpoint you run: a self-hosted vLLM, Ollama, or a fine-tune behind your gateway.
- Cloud on your key. Claude, OpenAI, Gemini, Kimi, and Perplexity on your own key, only when you pick one or place it in Stack. On your computer, the coding agent asks before a cloud model's first step in each task. An Ollama model whose tag (after the colon) is cloud or ends in -cloud runs on Ollama's cloud, not on your computer, and counts as cloud.
- Open weights. What runs on your own hardware is yours to control, and no app can take that back.
How the Stack works
One model runs the show and hands each task to the specialist that fits: coding, writing, analysis, image reading, or a fast model for quick turns. Anything without a specialist, it does itself. That model is the Reasoning model. You decide who sits in each seat, and you can change it any time.
- One planner. The Reasoning model routes; you never juggle models by hand.
- You place the specialists. Assign a model to a category, set when it is called, tune its effort.
- No lock-in. Swap any seat, keep the rest. Local and cloud sit side by side, and on your computer each seat says where it runs: Cloud, On this computer, or On your network.
Free and open source.