Friendly UI and server and agent for local LLMs

Run AI models on your own computer, and put them to work, with one download. Zest is a friendly interface for Lemonade, the open-source engine that runs language models locally, and for Pi, a coding-agent harness. The Zest AppImage includes both. Linux-only for now.

The idea is to provide a user interface that people who aren't necessarily steeped in technical know-how can still navigate to get open-weights models running on their own hardware, and then do real work with them: chat, draw diagrams, write documents, or let a model read and change the files in a folder while they watch. Download Zest, and you're good to go.

Chat, working in a folder: it read the code, fixed a bug with a one-line change (shown as a diff) and ran the tests again, step by step.

Contents: What you can do · Install and run · Finding your way around · Your first conversation · Choosing and managing models · Skills · Use your models in other apps · Keeping everything up to date · Advanced tools · Themes · Your data and privacy · For developers

Chat with an AI that runs on your own computer. Nothing you type leaves it, unless you let the model search the web: it asks first, every time.

Let it work on your files. Choose a folder and the model can read, search, edit and run things there while you watch. It asks before it changes anything.

Hand it a bigger job. A job keeps working in the background, round after round, and checks its own progress. You review what it changed before anything touches your folder.

Draw diagrams and write documents. Ask for a picture or a README and get one, shown right in the chat.

Teach it new tricks with skills. Short instruction sheets that make small models far more dependable. Add more by pasting a link.

Pick and manage models. Zest tells you which ones will run comfortably on your computer and downloads them for you.

Make it look the way you like. Fifteen themes, from soft plastic to Dracula and Nord, under Advanced > Themes.

Use your models in other apps. Zest can share them with tools such as opencode or any app that talks to an OpenAI-style server.

Zest runs on Linux. It needs no technical setup: you download one file and follow a few short steps. The Docs tab inside the app explains everything in plain language (its text lives in zest/docs.md).

Download Zest-<version>-x86_64.AppImage from the latest release.

chmod +x Zest--x86_64.AppImage ./Zest--x86_64.AppImage

Follow the welcome steps. Zest opens in your browser and walks you through four short steps in plain language, one screen at a time:

The welcome steps finish with a button that opens the chat.

People who already have models, or run their own Lemonade, can skip to Advanced > Setup guide, which keeps the full set of choices: where models are stored, models you already have (LM Studio, Ollama, llama.cpp and others are found for you), and Lemonade updates.

Linux, x86_64. The built-in Lemonade needs a reasonably recent system (glibc 2.38 or newer: Ubuntu 24.04, Debian 13, Fedora 39 or later). On an older system Zest tells you so and you can connect it to a Lemonade you install yourself instead.

A few common system libraries that almost every desktop Linux already has (OpenSSL 3, zstd, libdrm). If one is missing, Zest names it.

For GPU use, your graphics driver. Zest and Lemonade download everything else.

Internet access for the first-run downloads (the backend and your models).

If your system has no FUSE support, run ./Zest-*-x86_64.AppImage --appimage-extract-and-run instead. Closing the browser tab quits Zest (and its Lemonade) after about three minutes; launching it again while it's running just reopens the page.

The left-hand menu has six places. Everything a newcomer needs is in the first five; everything technical sits behind Advanced, and so does the choice of theme.

The menu on the left, with the Docs tab open.

Open Chat and type. The bar under the title shows which model will answer and whether it is in memory. If it isn't, Zest loads it when you send your first message, so that answer takes a little longer.

Chat runs on Pi, a coding-agent harness that ships inside Zest. Every chat is a co-working session: the model reads, searches, edits and runs commands (it has a terminal) in a folder while you watch. Pick the folder, or leave it and it works in your home folder, and says so at the top of the chat; there it starts on Ask first, and jobs need a project folder. It answers questions like a normal chat too, and asks before it changes anything (tool cards show each step; edits show a diff). Choose how careful it is per chat (Ask first, Edit freely, Do everything, Look only), add more folders it can look at or change, switch models, watch how much of the context window is used, and add notes while it works.

For bigger work, press Start a job…. Say what should get done and how Zest can tell it is done, such as a test command that passes. The job runs in the background, round after round, and stops when the check passes, when it hits its round or time limit, or when it stops making progress. By default it works on a copy of your folder; the Changes tab shows what it changed and you apply what you want. Jobs are listed under their chat, and you can keep chatting while they work. Chats and jobs are saved and pick up where they left off. See docs/COWORK.md.

Replies stream in as they are written. The bar under the title shows which model answers, how much context it has (128k by default) and whether it is in memory, with Load, Unload and Reload. A reasoning model's thinking appears live above its answer and folds away when it is done. Markdown and code blocks come with copy buttons, and a small strip shows what the model is doing: 📖 reading, 🤔 thinking, ✍️ writing, 🛠️ working.

Every chat works in a folder (your home folder unless you choose one) and can read, search, edit and run things there. Tool cards show each step, edits show a diff, and in Ask first and Edit freely modes it stops and asks before it runs a command. Here it was asked to fix a folder whose tests were failing: it read the tests and the code, made a one-line change and ran the tests again to confirm.

Lemonade handles models; Zest gives you the choices.

The Models page is for downloading models and editing how they run. My picks leads with the models I've tested and recommend; On this computer lists what you have. The meter on each card shows what will fit in the memory that is free on your graphics cards right now; if another program is using it, Zest warns you and names it.

Get from Hugging Face is for experts. It lists the GGUF versions in a repository, with a plain label, the size, and whether each fits. When Lemonade or llama.cpp refuses a model (for example because its architecture isn't supported yet, it needs a login, it can't be found, or memory or disk is short), Zest shows a plain-language message with the technical details tucked underneath, and offers to open Engines when an update may help. How do I choose a model? explains what to look for and links to benchmark sites and communities.

Press Settings on any downloaded card. Set the backend and devices, and tune llama-server with tick boxes and dropdowns instead of flags. Rows you have changed are highlighted, the exact arguments Zest will pass are shown, and anything your llama-server does not understand is flagged. Save keeps it; for a model that is in memory, press Reload in the chat's model bar to use the new settings.

Skills are short instruction sheets the model reads when a job calls for them. Zest comes with Superpowers (a trimmed copy of obra/superpowers: ask what you really want, plan, test first, find a bug's cause before fixing it, prove it works before saying so), a diagrams skill (the model writes Mermaid and Zest draws it in the chat, with Copy and Save as image; if it doesn't draw, one button asks the model to fix it) and a markdown-writing skill. To add more, open Skills, paste a link to a GitHub repository (or to one folder in it) and press Add skills. Any repository that follows the Agent Skills format works; Hermes Atlas lists many. Switch skills on and off one by one, or by set, and type / in a folder conversation to pick one. Only add skills from people you trust. See docs/COWORK.md.

Advanced > Server gives you an address, a key and copy-paste setup for tools such as opencode and Hermes, or any chat app, coding assistant or agent that can talk to an OpenAI-compatible server. Paste the address (http://127.0.0.1:13308/v1 by default), the key and a model name, and the other app uses the models on your computer.

Address, key and copy-paste setup for opencode, Hermes and any OpenAI-compatible app. The key is hidden until you press Show, including in the code samples. While an app is working you see a card with a glass that fills as it writes, a pill timeline of the last minute, and a request list. The Lemonade activity log is at the bottom of the same page.

The same page shows what those apps are doing:

Other apps never talk to Zest's own page. Zest opens a second, much smaller door on its own port (13308) that serves only the chat-style /v1 endpoints (models, chat, completions, embeddings, responses, audio, images, reranking, and loading or unloading a model). It always asks for the key, and pulling, deleting, installing or changing settings are not reachable through it. It listens on this computer only. Other devices on my network turns on listening on your local network; anyone there who has the key can then use your models, and the traffic is not encrypted, so use it on a network you trust. Make a new key invalidates the old one at once.

Advanced > Setup guide, where Lemonade and Pi are updated, models are brought in and backends are added.

Updating Zest itself is just downloading a newer AppImage. The two parts inside it, Lemonade and Pi, can be updated from inside Zest without a new download.

The AppImage ships with a recent Lemonade, but Lemonade moves quickly, so Zest can update its own copy without you downloading a new Zest.

Pi is the part that lets a model read files, make edits and run steps. It is updated separately, from Advanced > Setup guide, at the bottom, under The agent helper.

Advanced > Hardware: a tank for each graphics card, with its memory, load, temperature and power, and the switch and arrows that choose which cards Zest uses and in what order.

Already have Lemonade Server installed (for example as a system service)? On Advanced > Setup guide choose A Lemonade I already run, enter its address and, if needed, its API key. Zest will use it instead of starting its own, and won't touch the built-in copy. Switching back later is one click. Starting Zest with --lemonade HOST:PORT does the same without saving anything.

A few things work differently with a Lemonade you run yourself: Zest can't update it (use your package manager), and if it runs as a system service it may not be allowed to read model files in your home folder.

Zest looks the same everywhere you use it, but you can change its finish. Open Advanced > Themes and pick one. The choice is remembered on this computer, and nothing moves: only the colours and the surface change, so every button, panel and menu stays where it was.

Light Touch is the default: a calm cream page with a little hardware about it (keycap buttons, a pressed-in reply box and a small lamp that shows when a model is loaded). Two themes go further, and the rest dress Zest in colour schemes from popular code editors.

...and more! One Dark Pro, Dracula, Night Owl, SynthWave '84, Cobalt2, Nord, Gruvbox, GitHub Dark and Catppuccin are in the same picker.

What Where

Zest's settings, including the built-in Lemonade's API key and the key for other apps (file mode 600) ~/.config/zest/settings.json

Lemonade updates you've downloaded ~/.local/share/zest/lemonade/<version>/

The built-in Lemonade's own config and the backends it installs ~/.local/share/zest/lemonade-data/