Skip to content

Language models

This is the design, not what runs today

There is no model in this version: no self.model, no capability request, no cost accounting. Every step of the workflow works without one, which is what lets this wait — see the roadmap for what is funded and when.

One seam is already cut. A plugin that brings its own model can answer what is typed into the Prompt today, through podlibre_answer — it calls the model itself, and none of the resolving, installing or metering below is involved.

Podlibre completes every step of the workflow without a model. Where one is available it writes show notes, proposes chapters, drafts posts for social networks, and answers questions typed into the Prompt.

Ollama on the user's own machine is the recommended path and the default. Nothing leaves the computer, there is no account and no bill. External providers — Anthropic, OpenAI — are supported on the user's own key for those who want them.

What a plugin may ask for

Three ways, in descending order of politeness:

Request Example When to use it
A capability summarise, long_context Almost always. The user's existing model probably qualifies.
A recommended model "works best with qwen2.5:14b" When you have tested one and it is clearly better.
An exact model llama3.1:8b Only when nothing else will do.

Podlibre resolves the request against what the user has installed and offers to install what is missing. A plugin that names an exact model on a machine that cannot run it is a plugin the user cannot use, which is why asking for a capability is preferred.

Installing models

Podlibre lists the models a user has, the ones their installed plugins ask for, and what each would cost in disk space and memory. Installing is one button; Ollama does the work in a task like any other, with a progress bar and a Stop button.

What it costs

Every call is recorded against the plugin that made it: the model, tokens in, tokens out, and the cost where the provider charges for it. Preferences shows the month's spending per plugin.

Local calls cost nothing and are still counted, because knowing that a plugin is chatty matters even when it is free — it is the user's memory and battery either way.

A plugin using a model somewhere else can be held to a monthly ceiling in euros. Past it the answer is a refusal the plugin can handle, rather than a bill the user did not expect.

The ceiling is enforced where the calls are made -- by whatever provides model access -- and not by anything watching the plugin. Plugins share Podlibre's process and a plugin determined to call an outside service directly can; what this prevents is the ordinary case of a plugin costing more than its author realised.