The Basics

When you open Loca you land in a session: one conversation with the agent. You type in the bar at the bottom, and each reply arrives above it. That much is familiar from any chat app. This page covers the few things that are particular to Loca, the ones worth knowing before everything else falls into place: the two programs that make it run, the session you work in, and the ways to put more than plain text into a message.

Core and Client

Loca is built from two programs with a clear division of labor. The Core is a service that holds your data and does the heavy lifting: it stores sessions, runs models, executes tools, and reads your memories. The Client is the desktop app you see: it draws the conversation, captures what you type, plays audio, and otherwise stays thin. Most of this manual makes more sense once you hold that split in mind, because nearly every setting belongs to one side or the other.

Because the Core is a separate service, it does not have to live where the Client does. The two usually run on the same machine, but a Client can pair with a Core on a more powerful desktop and drive it from a laptop. The Core bar at the bottom of the app tells you which Core you are connected to and how it is doing.

The session you work in

A session is a single conversation, with its own history and its own context, and Loca saves every one locally so you can leave it and return later. The bar across the top carries the two things you keep an eye on: its title, and a running measure of the context window, how much of the model’s attention the conversation has used so far. Loca names a session for you from your first message, so you never stop to title it.

The context window is finite, and the bar’s fill is your warning that a long session is nearing the edge. The fix is rarely to fight a crowded context: when what you are asking now has drifted from how the session began, start a fresh one. A new session gives the model a clean window and your full prompt its undivided attention, and the old session stays saved if you need it. Keeping several short sessions, one per thread, beats piling everything into one ever-growing conversation.

Saying more than text

The input bar is a plain text box, but it is also where you pull in files and reach for things you have already saved, so it is worth knowing what it can do beyond typing.

• Attachments. Attach a file, drop in a screenshot, or capture a region of your screen, and it stages above the bar for review before you send. A document is passed along as text the agent can quote and work from; an image is only read if your model supports vision, so a text-only model cannot see one.
• The #, @, and / symbols. Type one of these and an autocomplete list opens for things you have saved: your snippets and memories. Pick a match and Loca drops its contents into your message without your hands leaving the keyboard.
• Dictation. If you would rather speak than type, the bar can show a microphone for Speech-to-Text, transcribing what you say into the same box.

How It Works

The two programs talk over a small local network connection: the Client sends requests to the Core and streams replies back over a live link, which is why a reply can keep arriving even if you switch away and come back. The Client holds no conversation state of its own; it renders what the Core reports.

Everything Loca stores lives under ~/.loca on the machine running the Core: your sessions, memories, snippets, downloaded models, and the settings you have changed. A session is a plain file on that disk, not a row in someone else’s database, which is what local-first means in practice. Nothing about a conversation leaves that machine unless you point a setting at an External provider, and Loca tells you when you have.

With the loop in hand, the next chapter is the engine behind every reply: Language Model.