Hand a card to a model on your own machine
An Ollama card does not shell out to a chat REPL. PortBay's own in-process runtime works the card against a model on your Mac — calling tools, editing files, checking the running app — with nothing leaving the machine.
Can an AI coding agent work a PortBay task offline on a local Ollama model?
Yes, and the mechanism is worth understanding because it is not what the name suggests. Assigning Ollama to a card does not run `ollama run` — that is a chat REPL with no tool loop, and PortBay explicitly does not dispatch it. Instead the card is routed to PortBay's own in-process runtime, which drives the local model through a real agent loop: it calls tools, edits files in the card's git worktree, and streams progress back onto the card. The model itself comes from the Ollama server PortBay installs and supervises, and the catalog you pick from is the set of models actually present on your Mac rather than a static list — the model you choose becomes the positional argument, not a flag. Because there is no provider on the other end, an Ollama card needs no API key and no account, and the entire run stays on the machine: it works with the network off. Headless and interactive both execute the same way; interactive additionally opens a `portbay tui` window onto the run.
A card worked entirely on-device
No provider call, no key, no network. The tool loop, the edits and the progress stream all run inside PortBay.
Your first fully local card
Five steps from a clean Mac to a card worked by a model on your own hardware. Every command runs as written.
- Install PortBay
- Let PortBay run the Ollama server
- Pull a model worth giving real work to
- Add your project and press play
- Assign Ollama to a card and move it to To Do
Install PortBay
Homebrew installs the app and the bundled portbay CLI. Prefer a direct download? Grab the DMG from GitHub Releases and drag PortBay.app into Applications.
brew tap portbay-app/portbay && brew install --cask portbayLet PortBay run the Ollama server
PortBay installs, supervises and updates a local Ollama server the way it supervises a project runtime, and exposes one on-device endpoint. If you already run your own, point PortBay's preference at that binary instead.
127.0.0.1:11434Pull a model worth giving real work to
The picker lists the models actually present on your Mac, not a catalog of what exists. A small model will happily fail a real task — pick one sized for the work and for the machine you are on.
ollama pull qwen2.5-coderAdd your project and press play
The project gets its runtime, its .test hostname with trusted HTTPS and its database. A local model needs a checkable environment more than a frontier one does, not less.
portbay add ~/code/myappAssign Ollama to a card and move it to To Do
PortBay's runtime works the card against your chosen model — tools, file edits, progress on the card. Pick interactive if you want a portbay tui window onto the run; it does not change what executes.
.portbay/tasks/tidy-migration-names.mdWhy `ollama run` is not this
The obvious command looks like the same thing and is not, which is the single most useful fact on this page.
Same model, different harness. The harness is the part that turns an answer into a change you can open in a browser.
ollama run qwen2.5-coder>>> refactor the invoice totals # a chat REPL: it can describe the change# it cannot open a file, cannot run a command,# cannot check the app — there is no tool loop # an Ollama card instead runs inside PortBay:# tools called, files edited, worktree branched,# progress streamed onto the card
The pieces underneath
Asked before downloading
Don’t see your question answered here? The docs go deeper — every feature, flag and failure mode.
Read the docsNo, and deliberately not. `ollama run` is a chat REPL with no tool loop — it cannot open a file or execute a command. An Ollama card is routed to PortBay's in-process runtime instead, which drives the model through a real agent loop that calls tools and edits files in the card's worktree.
Neither. There is no provider on the other end of a local-model card: the model runs on your Mac, through the Ollama server PortBay supervises. The run works with the network off.
The ones actually installed on your Mac. PortBay's picker reads the host's detected models rather than showing a static catalog, and the model you choose is passed as the positional argument that Ollama expects, not as a flag.
What executes is identical — both modes run through the same in-process runtime. Interactive additionally opens a `portbay tui` window onto the run so you can watch it. It changes what you see, not what happens.

Give your projects and your agents a real local home.
Download for macOSFree & open source · macOS 11+ on Apple Silicon · Pro from $10/mo