Desktop AI chat apps for Linux where you bring your own API key

The best desktop AI chat app for Linux where you bring your own key depends on what you want next to the chat window. Going by each project's own download pages, documentation and GitHub repositories, Chatbox, Jan, Cherry Studio and AnythingLLM all ship Linux builds, take API keys from OpenAI and Anthropic among other providers, and can work with local models too. Barside, the free desktop app we make, installs on Linux as an AppImage or a .deb and runs its assistant on the Claude, ChatGPT, Gemini or OpenRouter account you connect, right beside your notes and tasks.

We did not install or test the other apps for this guide. Every fact about them comes from the vendor's own pages as of October 16, 2026, linked where we use them, so open the page before you rely on a detail. Barside gets the same treatment, limits included: it is a notes app with an AI assistant built in, and it brings no AI model of its own.

Which desktop AI chat apps for Linux let you bring your own key?

Five desktop AI chat apps for Linux that let you bring your own key are compared below, and they differ most in packaging, local models and where the chats end up. Four of the five also work with local models, and Barside is the exception. The table sums up what each project's own pages say.

Chatbox

Chatbox's download page offers Linux builds for x86_64 and ARM64 and says "AppImage and deb packages are available; AppImage may require FUSE." Its Bring Your Own Key guide lists OpenAI, Anthropic (Claude), Google (Gemini) and OpenRouter, the local options Ollama and LM Studio, and a separate guide for DeepSeek, Kimi, MiniMax, GLM, ZenMux and others. The GitHub repository holds the Chatbox Community Edition, open source under GPLv3.

Jan

Jan's README on GitHub calls it "an open source alternative to ChatGPT that runs 100% offline on your computer." It downloads and runs open models such as Llama, Gemma and Qwen from Hugging Face, and it connects to cloud models from OpenAI, Anthropic, Mistral, Groq, MiniMax and others. Jan's Linux installation page says "Flatpak is the recommended way to install Jan on Linux," with .deb and AppImage files as the other routes. It lists a processor with AVX2, at least 8 GB of memory (16 GB recommended) and 10 GB of free space, and the README gives the license as Apache 2.0.

Cherry Studio

Cherry Studio's README describes "a desktop client that supports multiple LLM providers, available on Windows, Mac and Linux." It names OpenAI, Gemini and Anthropic among the cloud services, lists web service integration for Claude, Perplexity and Poe, and supports local models with Ollama and LM Studio. The latest release on GitHub, v2.1.6 when we looked, has AppImage, .deb and .rpm files for x64 and arm64. The Community Edition is under AGPL-3.0, and the README says commercial use of it is permitted in full compliance with that license.

AnythingLLM Desktop

AnythingLLM's Linux installation page says the desktop app "is packaged as an AppImage and is available for x64 and arm64 architectures." An installer script downloads it and offers to create the AppArmor rule that, per the page, the app needs to start. The LLM setup overview lists cloud providers such as OpenAI, Azure OpenAI, Anthropic, Google Gemini, OpenRouter, Groq and Mistral, and local ones led by a built-in provider, the default, then Ollama and LM Studio. In its own words, "We allow you to connect to both local and cloud-based LLMs - even at the same time!"

Does Chatbox work on Linux with your own API key?

Yes: the Chatbox desktop app for Linux works with your own API key, according to its download page and its guides. You pick the AppImage or the .deb for your processor, then add a key from the provider you use. Three details from Chatbox's own pages matter on Linux:

• Where requests go. The BYOK guide says "Data goes directly to AI providers, not through intermediary servers."
• Where chats and keys stay. The data storage FAQ puts chat history, settings, custom prompts and model settings in ~/.config/chatbox/ on Linux. It says chat history is not uploaded to Chatbox's servers, and that "Your API Keys are not sent to our servers."
• Moving to another computer. The same FAQ says "Chatbox currently does not support automatic multi-device sync." You export from Settings → Data Management → Export Data, as JSON for a full backup or Markdown for reading, and import the JSON file on the other device.

The BYOK guide sets your own key next to Chatbox's own AI subscription and says that with a key you "Only pay for Tokens actually used," while local models cost "No fees except electricity." The pages we read do not say whether Chatbox charges anything for the app itself when you only use your own key.

What does Barside offer on Linux, and what does it leave out?

Barside is a free desktop app for notes, tasks and capture with an AI assistant built in, and it runs on Windows, Linux and Mac. On Linux it comes as an AppImage for most 64-bit distributions, tested on Ubuntu, Debian, Fedora, Arch and openSUSE, or as a .deb for Ubuntu and Debian. The AppImage replaces itself when you update, and the .deb updates through the system password dialog.

Barside brings no AI model of its own. In Settings → AI you connect Claude, ChatGPT or Gemini with an API key, or OpenRouter with a one-click sign-in, and switch between connected companies with a click. On Linux the assistant, AI search, agents and Quick AI in the search bar all run on that connection. Barside's guides also describe My plan, which runs Claude Code or Codex on your computer, signed in with a Claude or ChatGPT plan you already pay for, so its limits apply instead of a bill per use. They do not say which systems My plan works on, and Anthropic and OpenAI have not said that apps may use a plan this way, so an API key is the safer choice.

The assistant works more like a helper than a plain chat window. It answers in a conversation, and it can also write and edit notes, add tasks and reminders, and save pages to Reading, with each step shown in an activity log. After each reply that changed something, an Undo button reverts it, and conversations are kept, so Undo still works after a restart.

What Barside leaves out, or does differently on Linux:

• Local models. Its guides describe no local model for the assistant, so every AI answer needs an internet connection and goes to the company you connected. If prompts must stay on your own hardware, one of the four apps above with a local model fits better.
• A long provider list. There are four companies to connect. OpenRouter passes your request on to the provider of the model you pick.
• Other apps. On Linux the assistant can list, focus and minimize windows and press keys, but it cannot read the controls inside other apps, and it collects only the current browser page instead of every tab in a window. Its window tools use X11, so on Wayland they see only apps running through XWayland.
• Hotkeys on Wayland. Global hotkeys work while an X11 app is active, and Barside gives you a command to add as a shortcut in your desktop's keyboard settings.
• Web search with Gemini. The assistant's web search works with Claude, ChatGPT and OpenRouter. Gemini has no web search in Barside.
• The tray on GNOME. GNOME needs the AppIndicator extension to show Barside in the tray.

Which Linux AI apps take OpenAI and Anthropic keys?

Every desktop AI app on Linux in this guide lets you bring your own API key from OpenAI and Anthropic. Chatbox, Jan, Cherry Studio and AnythingLLM name both companies on their own pages, and Barside connects them as ChatGPT and Claude. The differences lie in how the key gets in and where it stays.

• Chatbox says your API keys are not sent to its servers. The pages we read from Jan, Cherry Studio and AnythingLLM do not say where a key is stored.
• OpenRouter appears in Chatbox's guide, which files it under aggregators, in AnythingLLM's provider list and in Barside. It suits people who like to compare models from several companies with one account.
• Barside opens the key page of Claude, ChatGPT or Gemini in your browser, picks the new key up from the clipboard, checks it and saves it encrypted for your user account. It reads the clipboard for up to 10 minutes, ignores everything except that company's key, and removes the key from the clipboard. Keys go only to their own company and are left out of exported backups. OpenRouter is a sign-in with Authorize, with no key to copy.

What does bringing your own key mean for cost and data?

Bringing your own key means the AI company bills you directly and the app stays out of the AI bill. You create an API key in your account at OpenAI, Anthropic, Google or OpenRouter, add it to the app, and that company charges you for what each request uses. Light or uneven use often costs little, while long sessions and agents that run every day add up. A subscription app, by contrast, charges the same each month whether you use the AI or not.

For your data, the key decides who receives your messages, and the app decides where your history is kept. Every prompt you send to a cloud model reaches that AI company, and its data policy applies. Chatbox says its chat history stays on your device, Jan's Linux page names a local data folder, AnythingLLM's uninstall steps point to a local app data folder, and Barside keeps your library in its data folder. A local model keeps the prompt itself on your computer, at a cost in hardware: Jan's Linux page asks for at least 8 GB of memory and 10 GB of free space.

Two habits keep both in hand:

1. Set a monthly limit in your AI company's account, since its bill is the final word on what you spend.
2. Check where the app keeps chat history before you paste anything sensitive, and back that folder up with the rest of your home directory.

How do you connect your own key in Barside on Linux?

1. Download Barside and pick the AppImage, or the .deb on Ubuntu or Debian.
2. Open the assistant with Ctrl+Shift+A, or choose Assistant in the sidebar. The first time, it shows Connect an AI to start, with one button per company. The same buttons are in Settings → AI.
3. For Claude, ChatGPT or Gemini, Barside opens that company's key page in your browser. Make a new key and copy it, and Barside picks it up, checks it and saves it encrypted.
4. For OpenRouter, sign in and press Authorize.
5. Choose a model under the buttons. Claude and OpenRouter start with Claude Opus 5, ChatGPT with the newest balanced GPT model on your account, and Gemini with the newest Gemini Flash.
6. Under AI spending limit, turn on Limit what Barside's AI spends and set an amount from US$0.50 to US$10,000 for a day, a week (from Monday) or a month. Barside warns you at 80% and stops its AI at the limit.
7. Type in the box that says "Ask Barside to do something…" and press Run.

An example prompt to try first: "Create a note called Server migration with a checklist of these steps, and remind me about it tomorrow at 9:00." The note and the reminder appear right away, and Undo removes both if the result is not what you wanted. Each computer counts its own AI use against the limit, so set it on every computer where you use Barside.

Which app fits which kind of Linux user?

These are made-up examples, to show how the documented features line up with common needs.

• Lena, a data engineer on Fedora, wants her Anthropic key for hard questions and a local model on days she handles sensitive data. Jan lists Fedora among its distributions and runs models itself, and Cherry Studio publishes an .rpm and connects to Ollama.
• Marcus, a support lead on Ubuntu, wants one chat window for an OpenAI key and an OpenRouter account, plus a history he can export. Chatbox's .deb, its BYOK guide and its JSON or Markdown export cover that.
• Noor, a freelance translator on Debian, keeps client glossaries and deadlines in notes and wants the AI to work inside them. Barside's .deb with her Claude key puts the assistant next to her notes and task board, and she leaves one client's workspace out of Library AI in Library & briefing → Sources.
• Tomás, a market analyst on Arch, wants to switch between a local model and a cloud model in one app. AnythingLLM's docs say you can connect both, even at the same time.

Where do your chats, notes and key live in Barside?

Barside is local-first. Your notes, tasks and the assistant's memories live in Barside's data folder on your computer, conversations are kept by Barside, and you need no Barside account for any of it. When you use the assistant, your request, the tool results and memories it needs, matching excerpts from your library, and screenshots you ask about go to the AI company you connected, whose data policy applies. Workspaces you untick in Library & briefing → Sources stay out of Library answers and agents.

Your key is encrypted for your user account and sent only to its own company, with no Barside server in between for work on your computer. The exception is cloud agents in Barside Plus, which send your key to Barside's server, sealed, only to run your scheduled agents. Barside also counts which features are used: it sends barside.space the names of the features used that day, with the system and the Barside version, never your notes, files or who you are. It is on by default, and a switch in Settings turns it off. Read the AI you connect and your data for the full picture.

How does Barside compare with chat-first apps?

If you want a dedicated chat window with local models, the four apps above are built for that, and their docs show it. Barside fits when the chat is a means to an end: notes written, tasks added and pages saved, each with Undo. For more depth, read AI notes apps where you bring your own key, which also covers Raycast and Notion, and AI research assistant pricing, which sets subscriptions against paying for the API.

The AI note-taking app pricing guide works through costs in more detail, AI note apps and client confidentiality covers what reaches the AI provider, and the best local-first apps lists more software that keeps your work on your own device. For calls on Linux, see AI meeting notetakers that work on Linux.

Works well with