Side by side.
| Aspect | kibbu | Ollama on one machine |
|---|---|---|
| What serves a request | Whichever enrolled machine in your fleet is idle, plugged in and has the model. | The machine Ollama is installed on. |
| Who calls it | Your product, agents and jobs, from anywhere, with an API key. | Usually the person and the programs on that machine. |
| API | OpenAI-shaped. Change one URL, keep your client. | Its own REST API, plus OpenAI-compatible endpoints for a subset of the OpenAI API. |
| When the machine is busy | kibbu yields to the machine's owner and serves the request elsewhere in the fleet. | The request waits for, or competes with, whatever else that machine is doing. |
| When no machine can serve it | Falls back to OpenAI or Anthropic. You see every cent. | Set up in your application, or with a separate gateway. |
| Adding machines | A lightweight agent, installed with one click. Enrollment is token-based and explicit. | Install and run it on each machine, and point each caller at the one it should use. |
| Cost visibility | Every request logged, priced and attributed, local or cloud. | Nothing to meter: the work runs on your own machine. |
| Runs on | macOS, Windows and Linux. | macOS, Windows and Linux. |
Pick Ollama when.
- One person wants to run open models on their own computer.
- You are trying models out, offline or in development, and nothing else needs to call them.
- You want to manage the model runtime yourself, machine by machine.
Pick kibbu when.
- Your product, agents or nightly jobs call a model all day and you pay for every call.
- The company already owns laptops, workstations or servers that sit idle for part of the day.
- You want one endpoint for the whole fleet, with cloud behind it for what the fleet cannot serve.
Moving one workflow takes three changed fields. The quickstart shows them, and kibbu.md lists when kibbu is the wrong choice.
Questions.
Do I need to change my code to move from Ollama to kibbu?
If your code already uses an OpenAI client against Ollama's OpenAI-compatible endpoint, three fields change: the base URL, the API key and the model. The quickstart shows all three.
Does kibbu run on the same kind of hardware?
Yes. kibbu runs on macOS, Windows and Linux. Apple Silicon Macs and machines with NVIDIA GPUs are the ones that can serve the larger models.
Will it slow down the people using those machines?
kibbu works only on idle, plugged-in machines, respects battery and thermal state, and steps aside the moment the owner is back.
What happens when every machine is busy?
The request goes to OpenAI or Anthropic instead, priced, logged and attributed, and your application does not see the difference.
How is kibbu priced?
Pricing is agreed per pilot during early access.
