A local AI server, run like a project
One Start button, a supervised process, a live model catalog. PortBay manages Ollama the way it manages your dev servers, and everything in the app shares the one on-device endpoint.
How do I run a local LLM on my Mac?
Ollama is the usual answer: it downloads open-weight models and serves them over an HTTP API at 127.0.0.1:11434, the address every Ollama-aware editor and script already expects. Running it well is the fiddly part — ollama serve holds RAM until something stops it, picking a variant that fits 16 GB is guesswork, and the weights pile onto the boot disk. PortBay runs that server the way it runs your projects: one Start button and a supervised process. With no Ollama installed, one click downloads a signed build into PortBay's managed runtimes and keeps it updated; an existing Homebrew or Ollama.app install is detected and used instead. The Models view browses the live ollama.com catalog grouped by family — Qwen, Llama, DeepSeek, Gemma, Mistral, Phi — with size and freshness per variant and resumable pulls. Prompts stay on your Mac, with one badged exception: the cloud-served Kimi K2 family.
Verified against the product, 2026-07-10.
The playground on your own silicon
Chat, image generation, speech and embeddings run against the managed local server; the model list is the live ollama.com catalog.
From zero to a running model
- Get the binary
- Start the server
- Pull a model
- Run the smoke test
- Point everything at it
Get the binary
Click Download Ollama for a signed, managed build (with update checks), or let PortBay detect the install you already have from Homebrew or Ollama.app.
Start the server
A supervised ollama serve with your saved configuration. A server already running outside PortBay can be taken over; quitting PortBay stops the managed server so nothing keeps burning memory.
http://127.0.0.1:11434Pull a model
The catalog is fetched live from ollama.com and cached for offline use. qwen2.5:7b (~4.7 GB) is the recommended first pull for dictation rewrites, coding prompts and chat. Downloads survive navigation.
qwen2.5:7b · ~4.7 GBRun the smoke test
Test prompt verifies the endpoint, the selected model and the response path end to end, and reports round-trip latency. If another process owns the port, the diagnostic names it.
Point everything at it
Dictation rewrites, the SSH assistant and local-model card dispatch read the same endpoint, and so can your editor or scripts: it is a standard Ollama server.
Local AI, tier by tier
127.0.0.1:11434, the standard Ollama address; external tools keep working.
Model weights are relocatable: point the models directory at an external SSD and the AI page tracks that volume's free space.
Running Ollama as a hobby unto itself
The manual route works until you forget the server is eating memory, guess wrong on a model, or lose the port to something else.
PortBay supervises the server, curates the catalog with honest size guidance, relocates the weights, and shares one endpoint with everything.
$ brew install ollama$ ollama serve & # who stops this? it's holding RAM at 6 pm$ ollama pull ??? # which family? which size fits 16 GB?$ lsof -i :11434 # something else owns the port now# weights pile up on the boot disk; moving them is a symlink ritual# every tool (editor, dictation, scripts) configured one by one
One app, not a stack of tools
Dictation
Hold Fn anywhere on your Mac: on-device Whisper or Parakeet transcription plus a local AI rewrite, with ⌘Z always restoring your words.
Explore →Agent task board
A Kanban board both you and your agents work. Cards are Markdown in the repo; moving one to To Do dispatches the agent.
Explore →Agent dev environment
Stack detection, runtimes, HTTPS, databases and mail, provisioned per project — the running app your agents verify their work against.
Explore →Asked before downloading
Not for local models: inference runs on your machine against 127.0.0.1:11434 and the models directory you chose. The one exception is the Kimi K2 family, which is cloud-served by Ollama and requires an ollama.com sign-in; the catalog badges it 'cloud' so you cannot pick it by accident.
No. One click downloads a signed managed build into PortBay's runtimes folder, with update checks handled for you. If you already installed Ollama through Homebrew or the official app, PortBay detects it on PATH and common locations and uses that instead.
qwen2.5:7b, about 4.7 GB: the recommended default for dictation rewrites, coding prompts and local chat. The catalog groups families with one-line guidance and per-variant sizes so you can trade accuracy against your RAM honestly.
Yes. The endpoint is a standard Ollama server at 127.0.0.1:11434 (editable in Configuration), so anything that already speaks to Ollama keeps working unchanged, and PortBay's own consumers (dictation, SSH assistant, task dispatch) share it.
Anywhere you point the models directory; an external SSD is the usual choice since a handful of 7B models is 20 to 30 GB. Ollama models and speech-to-text models share one AI-models root, and the AI page shows the volume's free space next to them.
Give your projects and your agents a real local home.
Download for macOSFree & open source · macOS 11+ on Apple Silicon · Pro from $10/mo