Documentation

Getting started with LlamaBoss

LlamaBoss is a native Windows app for chatting with AI models. Run GGUF models on your own PC, add a hosted provider when you want one, and let the agent work with your files under rules you control.

Applies to LlamaBoss 0.1.29 for Windows and Mac. Updated November 12, 2026.

A chat with Agent Mode on: the model searches, reads part of a file, answers, then runs a test command after approval.
Documentation and installer versions

The current downloads are 0.1.29 for Windows and for Mac. This guide describes the 0.1.23 release; New in 0.1.29 and New in 0.1.27 list what changed since. It is written for Windows; Mac version lists what is different on a Mac. Notes marked From 0.1.23 describe what changed since 0.1.21. Version 0.1.22 was never released; its changes shipped in 0.1.23. Check your installed version with the ⓘ button and use Check for Updates to see what is available.

New in 0.1.21

You can connect through ChatGPT plan sign-in, add provider models by their exact ID, and export a conversation with usage and timing metrics. Conversation history and navigation have also been improved. See AI providers and Export with metrics.

New in 0.1.29

These changes arrived in 0.1.29, for Windows and Mac:

  • Your passwords stay private. Tools can no longer read stored credentials: git credential fill, gh auth token, saved Git and GitHub CLI tokens, LlamaBoss's own API keys and browser password files are refused, even when you have approved tools for the chat. On Windows this includes Credential Manager and DPAPI-protected secrets. On a Mac, the shell tool also refuses Keychain password reads.
  • Web pages open in your browser. When a model opens an .html file, it opens in your default browser instead of being read into the chat.
  • Local servers show their models. Add AI provider now lists the models of a running llama-server (the list used to come up empty), and the scan checks ports 8080 to 8090. See Add a provider.
  • Eject. Click Eject in Manage Models (right-click the model name at the top), or the eject icon on the Project/Skills strip, to unload the local model and free memory (VRAM on a PC with a graphics card). It loads again when you send your next message.
  • A steadier agent. Remote chats always get their tools, long chats keep their prompt cache, several read-only steps run together, and a model that says it will do something and then stops is nudged to do it. write creates missing folders, and UTF-16 text files read as text.
  • Excel and PDF fixes. Creating workbooks, inspecting and reporting on spreadsheets, CSV to Excel, reading PDF text and filling PDF forms all work more reliably.
  • Smaller things. The context meter no longer counts reasoning, provider settings and API keys are saved safely, LlamaBoss warns you before you rename or delete a key a provider uses, and notes_append adds to a notes file instead of overwriting it.

New in 0.1.27

These changes arrived in 0.1.27, for Windows and Mac:

  • One window for AI providers. Add and Edit AI provider is now a single screen: choose Local server, OpenRouter, OpenAI, ChatGPT plan or Custom across the top, connect on the left and pick models on the right. See Add a provider.
  • Finds local servers. Local server scans the usual ports on your computer and lists what answers, such as LM Studio, Ollama or another llama-server, with no key needed.
  • Context window per model. Each model can have its own context limit, read from the server when it reports one, so the context meter matches what the server accepts.
  • Faster photos. Large images are scaled down before they are sent, and phone photos arrive the right way up.
  • Retro Hi-Fi theme. A cassette-deck look, with cyan for actions and amber for code and readouts. See Appearance.
  • Clearer Copy. The code-block [ Copy ] button is bold and uses the theme's accent colour, and only the block you copied last says "Copied".
  • Tidier chats. Clicking + no longer leaves an empty saved chat behind, saved API keys are shown masked with a Replace key button, and the context meter no longer clips when you switch themes.
  • Images with MTP models. Models with built-in multi-token prediction that also read images keep image processing on the GPU, so reading an image takes seconds instead of minutes.

New in 0.1.23

These changes arrived in 0.1.23:

  • Notes as a folder. Notes become small files in %USERPROFILE%\LlamaBoss\Notes, one subject per file, and the model sees only a one-line index of them. See Notes.
  • Keys for Skill scripts. A script receives only the service keys its own SKILL.md or WORKFLOW.md declares. See Skills that use API keys.
  • Safer Skill import. A loose SKILL.md in a folder of unrelated files can be imported on its own.
  • Copy and hash tools. The agent can copy a file byte for byte and check SHA-256 hashes without writing a script. Fetched web pages report their size and SHA-256. See Agent tools.
  • Answers. Images in an answer can show inline, and a long hash in an answer that doesn't match any tool output is flagged. See Images and hashes in answers.
  • Calmer activity line. It appears only for tools that run longer than about 0.75 seconds.

Install

  1. Download the installer from the LlamaBoss home page.
  2. Run the .exe and finish the installer.
  3. Open LlamaBoss from the Start menu or the desktop shortcut.
Windows SmartScreen

The installer is code-signed by a verified publisher, so SmartScreen shouldn't warn you. If it ever does and you downloaded it from llamaboss.com, choose More info, then Run anyway. The full source is on GitHub.

Pick your first model

On first launch with no model installed, LlamaBoss opens the model downloader. You can also click the model pill in the top bar, or open Settings → Model, to download a curated model or paste a direct GGUF download link in Manage models, choose a GGUF already in your models folder, point LlamaBoss at a different folder, or pick a model from a provider you've connected.

Start small. A small model loads quickly and confirms your CPU or GPU runtime works before you try something larger.

Start a chat

  • Type a message and press Enter. Use Shift+Enter for a new line.
  • Add files with the paperclip, by dragging them onto the window, or by pasting.
  • Click the robot button to turn Agent Mode on or off for this chat.
  • Click · Auto next to the model name to change how much the model thinks before answering.
A good first test

Ask the model to explain a small text file or a screenshot. Then turn on Agent Mode and ask it to list its workspace. That checks plain chat, attachments, and a read-only tool in about a minute.

Explore the features

Local first, not local only

With a local GGUF model, your conversation is processed on your PC and works offline once the model is downloaded. LlamaBoss only goes online for features you use: model downloads, update checks and installer downloads, AI providers you connect, web pages the agent fetches, and Python packages you approve. Code run by the agent can also access the network.

Privacy & data lists every storage location and network feature.

Mac version

LlamaBoss for Mac runs on Apple silicon Macs (M1 or later) with macOS 13.3 or later. The current Mac version is 0.1.29. Download it from the home page, open the disk image and drag LlamaBoss to Applications. It is signed and notarized by Apple, so it opens without a warning.

It has the same features, agent and approval rules as the Windows version, adapted to the Mac:

  • Local models run on llama.cpp with Metal. Download Models states each model’s needs as the Mac memory size to look for.
  • Shell commands run in zsh, the Mac’s own shell, instead of PowerShell. The agent’s shell tool follows the same read-only and approval rules.
  • Python tools use the Python 3 that comes with LlamaBoss (from 0.1.24), so there is nothing to install. PDF text and forms, Excel and Word helpers work offline. Extra packages you approve go into LlamaBoss’s own Python environment, so any Python on your Mac is not changed.
  • Saved API keys are encrypted with a key kept in your login Keychain.
  • Settings is in the LlamaBoss menu (⌘,). File → New Window, or the Dock icon’s menu, opens another window.
  • Files: app data, settings, logs and models are in ~/Library/Application Support/LlamaBoss; chats, Projects, Skills and notes are in ~/LlamaBoss.

Not available on the Mac yet: ChatGPT sign-in, and in-app Download and Install updates (Check for Updates tells you when a new Mac version is out and opens the download page). The rest of this guide is written for Windows; where it mentions PowerShell or Windows paths, use the Mac equivalents above.