When PewDiePie pushed Odysseus to GitHub in late May, the repo cleared 30,000 stars in two days. The interesting part was not the celebrity bump but what the project actually was: a ChatGPT-style web app you run on your own machine, pointed at whichever model you like. That category, the self-hosted AI workspace, has been quietly maturing for two years. The good ones now match the polish of the hosted chat apps while keeping your conversations, files, and search history on hardware you control.
We tested 7 of the strongest self-hosted AI workspaces for desktop. Most run in Docker or a browser and work on Windows, macOS, and Linux without code changes. The benchmark was the boring stuff: how fast a fresh container becomes a usable chat window, whether file upload and web search work without extra plumbing, how the app behaves when two people share one server, and how much it lags ChatGPT on the day a new model lands.
What to look for in a self-hosted AI workspace
A handful of criteria separate the workspaces that survive a month of daily use from the ones that get docker-pruned:
- Provider agnosticism. The best apps speak OpenAI, Anthropic, Google, Mistral, Ollama, and any OpenAI-compatible endpoint from one settings panel. Tying yourself to a single backend defeats the point of self-hosting.
- Conversation organisation. Folders, tags, search across history, branching, and pinning. ChatGPT got this right late; the open-source clones now ship it on day one.
- Multi-user mode. Per-account chat history, role-based access, and quotas. A household or a small team needs this; a solo dev does not.
- RAG and file upload. Dropping a PDF or a folder into a chat and getting grounded answers. The serious workspaces ship a built-in vector store and a chunking pipeline.
- Web search and tool calls. A live search integration so the model can answer questions about this week, plus an MCP or plugin layer for tool use.
- A UI that holds up on a phone. Most of these are PWAs. If the chat window breaks on a 390px screen, you cannot use the home-lab server from the couch.
Quick comparison
| App | Best for | Platforms | Free plan | Standout feature |
|---|---|---|---|---|
| Open WebUI | The default self-hosted ChatGPT clone | Docker (any OS) | Yes (open source) | Multi-user accounts, RAG, web search, and a model marketplace |
| LibreChat | Power users who want every provider in one window | Docker (any OS) | Yes (open source) | Native MCP support and conversation branching |
| Lobe Chat | A polished UI that feels closer to the hosted chat apps | Docker, Vercel, browser | Yes (open source) | Agent groups that collaborate in parallel |
| AnythingLLM | Workspace-style RAG over private documents | Windows, macOS, Linux, Docker | Yes (open source) | Per-workspace document grounding with a no-code agent builder |
| Big-AGI | A research playground with personas and beam search | Browser, Vercel, Docker | Yes (open source) | Side-by-side model comparison and a developer-friendly chat surface |
| Odysseus | A local-first workspace with chat, agents, research, and docs in one app | Windows, macOS, Linux | Yes (open source) | Bundled deep research, document editor, and email in one self-hosted bundle |
| Hollama | A zero-fuss front end for an Ollama server | Browser, Windows, macOS, Linux | Yes (open source) | A single static page that talks to Ollama with no backend of its own |
The 7 best self-hosted AI workspaces on desktop
1. Open WebUI — best default self-hosted ChatGPT clone
Open WebUI is the project most people land on when they search “self-hosted ChatGPT”. One Docker command, a web app at localhost:3000, and a UI that mirrors the ChatGPT layout closely enough that a new user does not need a tour. It connects to Ollama out of the box and to any OpenAI-compatible backend with a base URL and a key. The latest builds bundle a model marketplace, prompt and tool libraries, RAG with per-document permissions, and a web-search switch that wires up SearXNG or Google PSE.
Where it falls short: Multi-user setup and SSO need a bit of reading. The interface is dense once you turn on every feature, and the mobile PWA is good rather than great.
Pricing:
- Free: open source, no licence fee
- Paid: optional enterprise tier with SLA and SSO support
Platforms: Docker, accessible from any modern browser on Windows, macOS, or Linux
Download: openwebui.com
Bottom line: Pick Open WebUI for a self-hosted AI workspace if you want the ChatGPT experience on your own server with the largest community behind it.
2. LibreChat — best for power users who want every provider in one window
LibreChat sets itself apart by being aggressively provider-agnostic. OpenAI, Anthropic, Google, Mistral, AWS Bedrock, Azure, and any OpenAI-compatible endpoint sit in the same model picker. Conversation branching lets you fork a chat at any message and explore two paths in parallel, which is genuinely useful for drafting work. Recent releases ship native MCP support, an agents framework, OAuth, LDAP, OIDC, a RAG API, and Meilisearch for full-text search across history.
Where it falls short: The compose stack is heavier than Open WebUI, with MongoDB, Meilisearch, and a RAG sidecar to manage. First-run configuration takes longer.
Pricing:
- Free: open source, no licence fee
- Paid: none
Platforms: Docker, accessible from any modern browser on Windows, macOS, or Linux
Download: librechat.ai
Bottom line: Pick LibreChat for a self-hosted AI workspace if you juggle several providers and want branching, tools, and search history in one window.
3. Lobe Chat — best polished UI that feels closer to the hosted apps
Lobe Chat is the most visually refined of the open-source workspaces. Smooth animations, careful typography, a plugin marketplace, and an Agent Groups feature where several agents collaborate on a task in parallel. It runs as a Next.js app you can deploy to Vercel for free or run in Docker, and it speaks to most major providers plus Ollama and any OpenAI-compatible endpoint.
Where it falls short: Some advanced workspace features sit behind a cloud-hosted Pro tier. RAG and document upload are present but feel less central than they do in AnythingLLM or Open WebUI.
Pricing:
- Free: open source, no licence fee
- Paid: optional Lobe Chat Cloud plan for hosted deployments
Platforms: Docker, Vercel, accessible from any modern browser on Windows, macOS, or Linux
Download: lobehub.com
Bottom line: Pick Lobe Chat for a self-hosted AI workspace if visual polish and agent collaboration matter more than the deepest RAG pipeline.
4. AnythingLLM — best for workspace-style RAG over private documents
AnythingLLM treats the conversation as a workspace bound to a set of documents. Upload PDFs, point at a website, connect a Confluence space or a GitHub repo, and every chat in that workspace is grounded in those sources. The agent builder lets non-technical users wire up web browsing, code execution, and custom skills without writing code. The desktop build is a real installer with no Docker required, which makes it the easiest entry point for someone migrating from ChatGPT Plus.
Where it falls short: The desktop app is single-user; multi-user mode needs the Docker server build. The chat UI is functional rather than beautiful next to Lobe Chat.
Pricing:
- Free: open source, no licence fee
- Paid: optional hosted Cloud plan and a Pro tier for advanced features
Platforms: Windows, macOS, Linux native installer, plus Docker for multi-user
Download: anythingllm.com
Bottom line: Pick AnythingLLM for a self-hosted AI workspace if your day involves talking to your own documents and you want a one-click desktop install.
5. Big-AGI — best research playground with persona and beam features
Big-AGI leans into the playground angle. Side-by-side model comparison, beam mode for generating several candidate answers from one prompt, persona presets, a code execution sandbox, and a UI that exposes more knobs than most of this list. It runs as a static Next.js app you can deploy to Vercel, Cloudflare Pages, or a container, and the same build works as a PWA on a phone.
Where it falls short: Multi-user mode is intentionally absent; it is built for an individual or a shared single-user instance. RAG is lighter than in Open WebUI or AnythingLLM.
Pricing:
- Free: open source, no licence fee
- Paid: none
Platforms: Browser PWA, Docker, Vercel, accessible from Windows, macOS, or Linux
Download: big-agi.com
Bottom line: Pick Big-AGI for a self-hosted AI workspace if you spend your sessions comparing models and tweaking prompts rather than chatting with documents.
6. Odysseus — best all-in-one local-first workspace
Odysseus is the project Felix Kjellberg dropped on GitHub in May, and it is the only entry on this list that bundles chat, agents, deep research, a document editor, IMAP email, notes, tasks, and a calendar into one self-hosted app. The chat surface speaks to Ollama, llama.cpp, vLLM, and OpenAI-compatible endpoints. Deep research runs a multi-step web pipeline and writes a cited report. The document editor takes inline AI edits, and skills plus MCP cover tool use.
Where it falls short: The project is months old, so polish and edge-case handling lag the older workspaces. The breadth of the bundle means the surface area is larger than a focused chat app.
Pricing:
- Free: open source, no licence fee
- Paid: none
Platforms: Windows, macOS, Linux
Download: github.com/pewdiepie-archdaemon/odysseus
Bottom line: Pick Odysseus for a self-hosted AI workspace if you want chat, research, docs, and email under one local-first roof and you do not mind a young project.
7. Hollama — best zero-fuss front end for an Ollama server
Hollama is the minimalist on this list. A single static web page that talks to an Ollama server (or any OpenAI-compatible endpoint) with no backend of its own. Chats live in browser storage. The whole app weighs less than most workspace splash screens. It is the right pick for someone who already has Ollama running and just wants a clean chat window with sessions and a system prompt, without managing a database or a container stack.
Where it falls short: No accounts, no server-side history, no RAG, no plugin system. By design.
Pricing:
- Free: open source, no licence fee
- Paid: none
Platforms: Browser, plus desktop builds for Windows, macOS, and Linux
Download: github.com/fmaclen/hollama
Bottom line: Pick Hollama for a self-hosted AI workspace if you want a chat front end for your Ollama server with zero ceremony and no extra services to run.
How to pick the right one
If you want the safest default with the largest community, run Open WebUI.
If you bounce between OpenAI, Anthropic, and a local model and want branching and search, run LibreChat.
If visual polish matters and you want agents that collaborate, run Lobe Chat.
If your real workflow is asking questions about your own documents, install AnythingLLM.
If you spend your day comparing prompts and models, deploy Big-AGI.
If you want chat, deep research, a doc editor, and email in one self-hosted app, try Odysseus.
If you already have Ollama running and want a clean chat window over it, drop in Hollama.
FAQ
What is a self-hosted AI workspace?
A web or desktop app that gives you the ChatGPT experience (chat, history, file upload, search, agents) while running on hardware you control. It connects to whichever model provider you choose, including local models on the same machine.
Do I need a GPU to run a self-hosted AI workspace?
Not for the workspace itself. The workspace is a thin client. If you point it at a hosted provider (OpenAI, Anthropic), a small VM or even a laptop is enough. You only need a GPU on the model server, and only if you want to run the model locally.
Which self-hosted AI workspace is the most private?
All of them are more private than a hosted chat app, since your conversations live on your own server. Odysseus, Open WebUI, and AnythingLLM are explicit about local-first defaults and no telemetry. If you also point them at a local model server like Ollama, no chat data leaves your machine.
Can I use a self-hosted AI workspace with a phone?
Yes. Most of these run as a web app, so you open the same URL from your phone browser. Open WebUI, LibreChat, Lobe Chat, and Big-AGI all install as PWAs for a near-native feel. The server has to be reachable from your phone, which usually means running it on your home network or behind a Tailscale or Cloudflare tunnel.
