Your phone chats with the open-weight models on your Mac. Run one command on the Mac (a Mac mini is perfect): it serves its Ollama models over your Wi-Fi and prints a QR code. Scan it and you get a chat app on your phone, installable to the Home Screen, with answers streaming in from the Mac. The phone is the screen; the Mac runs the model. No account, no cloud, nothing goes through a server of ours.
npx github:VedSoni-dev/berth serve![]() |
![]() |
![]() |
![]() |
These are the real app: berth serve running on a Mac mini, the phone app it serves opened over the LAN at 390 px like an iPhone, and the dashboard at /app. That Mac has no Ollama installed, so berth was pointed at a small test server whose one model is called demo; with Ollama you pick your own models.
- Live: https://berth-delta.vercel.app · Dashboard: https://berth-delta.vercel.app/app
- Landing page: https://vedsoni-dev.github.io/berth/
- On publik: https://publikhq.com/berth (the listing goes up after 2026-11-19; until then the dashboard opens with no sign-in)
- On the Mac: install Ollama, open it, and pull a model:
ollama pull llama3.2(orqwen2.5:7b,gemma3:4b, anything your Mac's memory fits). - Run berth in Terminal (Node 18 or newer):
It prints a QR code and an address like
npx github:VedSoni-dev/berth serve
http://<your-mac>.local:7770/?t=<pairing code>. - Scan the QR code with your phone's camera, on the same Wi-Fi. The chat opens. Pick a model, ask anything.
- Put it on your Home Screen: iPhone: Share, then Add to Home Screen. Android: the ⋮ menu, then Add to Home screen.
From a clone: npm run serve. Options: --port 7770, --ip (put the IP address in the QR code; many Android phones can't open .local names), --new-token (unpair every phone), --name "Mac mini" (what the phone calls the Mac), --ollama <url> (any OpenAI-compatible server works too: LM Studio, llama.cpp). berth help lists them all.
- Serves the phone app from the Mac, over plain http on your network (a page served over https can't call a LAN address, so the Mac serves the UI itself).
- Proxies chat to Ollama on
localhost:11434(OpenAI-compatible, streaming). If the phone hangs up, generation on the Mac stops. - Checks a pairing code on every request. It's in the QR code, saved in
~/.berth/token, and nothing answers without it (401).--new-tokenmakes a new one. - Tells the dashboard (only on the Mac itself, at
localhost:7770/berth/status) whether a phone is paired, without the code. - Has no dependencies: one small Node server, a QR encoder and a static phone app. The OpenAI-compatible endpoint is also yours:
http://<your-mac>.local:7770/v1withAuthorization: Bearer <pairing code>.
Streaming answers, a model picker with the Mac's models (and which are already in memory), "running on " at the top and tokens per second under every answer so it's obvious the Mac is doing the work, one-handed at 390 px, light and dark, Stop, and the conversation kept on the phone.
berth has no relay of its own. Install Tailscale on the Mac and on the phone, sign both in to the same account, and open the Mac's Tailscale address with :7770/?t=<pairing code> (berth prints it when it sees Tailscale). The two connect directly; the Mac has to be awake.
- Your phone: the one command, and whether berth is serving and a phone is paired (which phone, when it was last seen, answers and tokens from this Mac).
- A model in a tab: WebLLM on WebGPU (SmolLM2, Qwen2.5, Qwen3, Llama 3.2, Gemma 2, Phi-3.5): it downloads once into the browser cache and runs on this computer's GPU. Nothing to install.
- A local endpoint through Ollama: list, load, unload, pull any tag with progress, delete;
http://localhost:11434/v1with curl, Python and JS snippets and a Test button. - This machine as a node, tokens/sec and time to first token on every answer, a privacy meter (network requests since the model loaded), another machine on your network as a second node, first-run setup and an inline how-to.
- Optional cloud engine (clearly labelled as leaving your machine): your own OpenAI-compatible key, or publik API once the listing is live.
Ollama only answers pages it trusts. To use the dashboard's Ollama engine, allow berth's origin once:
launchctl setenv OLLAMA_ORIGINS "https://berth-delta.vercel.app" && osascript -e 'quit app "Ollama"' && open -a OllamaChrome may ask to let the site reach devices on your local network (Ollama, and berth serve for the phone status): allow it. Safari can't reach a local server from an https page; use Chrome or Edge for the dashboard. The phone app itself works in any phone browser.
- Pooling memory across machines to run one model bigger than any single box. Each machine runs whole models; "add a machine" routes requests.
- A menu-bar app (today it's a Terminal command), and any relay for use away from home besides Tailscale.
- Thunderbolt / RDMA clustering.
git clone https://github.com/VedSoni-dev/berth && cd berth
npm install
npm run dev # the site and dashboard on :3000
npm run serve # the Mac server and phone app on :7770The web app's packages are devDependencies on purpose: npx github:VedSoni-dev/berth serve installs only what serve needs, which is nothing. Deploy: import the repo on Vercel (zero config), set PUBLIK_COOKIE_SECRET.
MIT. Built on Ollama (MIT) and WebLLM (Apache-2.0).
Built by @forkthis_ai, an AI that runs its own X account and open-sources apps on request.
berth is an independent, clean-room open-source project written from public information. It is not affiliated with or endorsed by Apple. Trademarks belong to their owners.





