Auto model router
Picks the smallest model that can handle each question — preferring ones already loaded — so you don’t have to think about it.
Private AI chat you actually control.
One workspace for open models on your hardware — including one that runs right inside the Mac app — and frontier models when you need them. Full history, accounts, and control stay on your machine.
curl -fsSL https://geekaholic.gitlab.io/omni-chat/install-macos.sh | bash- macOS
- Linux
- Docker
- iPhone & iPad soon
Free and open source under Apache 2.0.
What’s new
Recent releases removed most of the setup — and taught the app some new tricks.
Models show up on their own
The desktop app finds OpenAI-compatible servers — Ollama, LM Studio, llama.cpp, vLLM — on your computer and home network, and their models appear in the picker. Zero config.
A model built into the Mac app
Download Qwen 3.5 once and chat with no server at all. It runs on Apple Silicon through MLX — the 4B on Macs with 16 GB, the 2B on 8 GB.
Sign in with ChatGPT
Already pay for ChatGPT? Sign in once and use OpenAI’s Codex models for chat and image generation through your subscription — no API key to create or paste.
Coding agents that follow a plan
Ask for something big and the assistant drafts small tasks with checks. Approve once, then watch each task run and pass.
Chatting in three steps
No config files, no API keys required. Most people are chatting within a few minutes.
-
Install with one command
Paste one line into Terminal on a Mac or Linux desktop. It downloads, verifies, and opens the app. Get the command
-
Your models are already there
Models served on your computer or home network appear in the picker automatically. Or download the on-device model on a Mac, or sign in with ChatGPT. Mix them any time.
-
Chat — it stays yours
History, projects, accounts, and memory are stored in a database on your machine — not in someone else’s cloud.
AI Servers
Discover and manage servers
- This computerOllama4 models
- mac-mini.localLM Studio3 models
- nas.localllama.cpp2 models
- gpu-box.localvLLM1 model
10 models ready in the picker
Local first, not local only
Open models are good enough for most daily work. Daily work stays private and free on your hardware — you only pay, and send data, when the task needs frontier capability.
-
Your chats leave the buildingSensitive drafts and family questions sit on someone else’s servers.
Your chats stay yours
Full history lives in a local SQLite database. Local models keep everything on your network. Off-the-record mode skips saving when you want it.
-
Metered thinkingRate limits and token bills punish everyday use.
Start simple, grow when ready
Start with Ollama or LM Studio. Add frontier providers or extra machines later — same dropdown.
-
Local UIs stop shortMany Ollama frontends are single-user chat shells.
One place for all your models
One picker plus a smart router. Local and frontier models side by side — pick yourself or let routing choose.
Where do your answers come from?
- Runs on
- The on-device model inside the Mac app, or Ollama and LM Studio on the same computer
- Your prompt goes
- Nowhere — it stays on this computer
- Costs
- Nothing per message
- Great for
- Drafts, rewrites, summaries, and private questions
- Runs on
- Ollama, LM Studio, llama.cpp, or vLLM on another machine at home — found automatically
- Your prompt goes
- Across your home network, nowhere else
- Costs
- Nothing per message — it’s your hardware
- Great for
- Bigger models, coding help, brainstorming, and the whole family at once
- Runs on
- OpenAI, Claude, Grok, and others with an API key — or your ChatGPT plan through Sign in with ChatGPT
- Your prompt goes
- To the provider you picked, only when you pick it
- Costs
- Per token, or your existing subscription
- Great for
- Hard reasoning, deep research, and tough coding
History is stored on your machine whichever you pick. Or choose Auto and oMNI picks the smallest model that can handle each question.
Which one for what?
- Draft emails & rewrite textLocal is excellent
- Summarize notes, articles, docsLocal is excellent; frontier is usually overkill
- Private or sensitive chatsLocal is best — it stays on your machine
- Brainstorm ideas & plansLocal is great; frontier is optional
- Explain code errors or stack tracesLocal is great; use frontier if you’re stuck
- Hard reasoning or deep researchFrontier is stronger; local handles many cases
- Coding (writing & debugging)Frontier is much stronger; strong local models do well
Connect hosted providers (OpenAI, Claude, Grok, etc.) in the same dropdown whenever you want. Your chats stay in your database either way.
Built for real work, not just chatting
Screenshots from the actual app, with real model output — the parts that set oMNI Chat apart.
-
Projects, tags, and live code previews
Organize history with projects and hashtags. Ask for a small app, chart, or SVG and open a sandboxed live preview right next to the conversation.
-
A private model that lives in the Mac app
Turn on the on-device model in Settings → Models and the app downloads Qwen 3.5, then runs it on your Mac’s GPU with MLX. No server to install, and once it’s downloaded it works offline — with thinking, web search, and tools.
Apple Silicon on macOS 26 or later. Qwen 3.5 4B (about 3 GB) on Macs with 16 GB of memory; the 2B on 8 GB Macs.
-
Zero config: it finds your models
Any OpenAI-compatible server on your computer or home network — Ollama, LM Studio, llama.cpp, vLLM, and more — is detected automatically, and its models simply appear in the picker. No URLs to type, no files to edit. Scan now in Settings → AI Servers digs deeper, including servers that need a key, and every connection can be tested before you save.
Network discovery is on in the Mac and Linux apps; Docker and server installs look on their own machine by default. Admins get the same Settings in the web version.
-
Use the ChatGPT plan you already pay for
In Settings → Integrations, Sign in with ChatGPT connects OpenAI’s Codex models through your subscription — for chat and for image generation. No API key, no per-token bill. The owner signs in once and everyone on the instance can use it, with history still stored on your machine.
Needs a ChatGPT plan that includes Codex. Prompts you send to these models go to OpenAI, like any hosted model.
-
Hand big jobs to coding agents — with a plan
For anything with several parts, the assistant writes a plan: a goal and small tasks, each with checks it can run. Edit or approve it once. A job card shows every task’s status, files, and check results live, and a failed task pauses for a fix instead of plowing on.
Runs on a coding agent such as Pi or Hola, in a workspace folder you can open and watch.
-
Run code without leaving the chat
When a question needs real math or data, the model writes Python or JavaScript and runs it in a sandbox inside your browser — nothing executes on the server. You see the code and its output, and the answer is built on the real result.
-
Memories you can inspect and pin
When a chat teaches the assistant something about you, a Remembered card appears with Undo. Every memory is visible, editable, and pin-able — you decide what it keeps.
-
Images, tools, and multi-user accounts
Generate images, call tools, and share one instance with family or teammates — each with their own account and history on the models you configured.
Image generation runs on a separate local image model — a capable GPU helps.
One install. Everyone gets their own.
Run oMNI Chat on one computer at home and give each person an account. They open it in a browser on their laptop, tablet, or phone — nothing to install on their side, and no subscription per person.
-
A private space each
Every person’s chats, projects, and memories are theirs alone. Other accounts can’t see them — or even tell they exist.
-
You stay in charge
Add family members in Settings → Members, switch an account off, and choose who else can manage the AI servers.
-
Set up once, shared by all
The models you connect — and a single ChatGPT sign-in or API key — work for everyone. No per-person plans or keys to juggle.
-
On every screen
Laptops, tablets, and phones just use the browser — and on iPhone, Add to Home Screen makes it feel like an app.
Sharing with other devices uses the Docker or home-server install. On the Mac and Linux apps, family members can still take turns on that computer, each signing in to their own account.
iPhone & iPad apps
Coming soonNative iPhone and iPad apps are in testing now. On iPad, the on-device model runs right on the tablet — no server needed.
Can’t wait? If oMNI Chat runs on Docker or a home server your phone can reach, open it in Safari, tap Share, then Add to Home Screen. It opens full screen with its own icon, just like an app.
And the rest of the toolkit
Math that renders
Equations display properly, in light or dark mode.
MCP tools you approve
Plug in Model Context Protocol servers for extra tools. Each person can switch servers on or off, and tools marked sensitive ask before every call.
Web search, one click
In the desktop app, start a private SearXNG from Settings → Local services; on Docker it’s one overlay. Search is a per-chat switch.
Fetch any URL
Point a model at a web page or YouTube video and it pulls in the readable text or transcript.
Off the record
Need a chat that isn’t saved or remembered? Start it off the record and it never lands in your history.
Works with your runners
Ollama, LM Studio, llama.cpp, vLLM, LiteLLM, and anything OpenAI-compatible — found automatically on your network — plus hosted providers.
Export chats
Download Markdown, HTML, or PDF when you need work to leave the app.
Attachments
Drop in images, PDFs, and text documents. Images go to vision models; document text is read in as context.
Voice in & out
Dictate with speech-to-text and hear replies read aloud, with optional local voice models.
Tidy thinking
A model’s reasoning folds away behind a live timer, and a guard stops models that get stuck in a loop.
Mac, Linux, or Docker
Native apps for Apple Silicon Macs and x86_64 or ARM64 Linux. Docker Compose for servers; iPhone, iPad, and Windows are next.
…and more. Read the full Manual
Get oMNI Chat
Install the Mac or Linux desktop app in one command, or run the Docker stack on a server for the whole household.
curl -fsSL https://geekaholic.gitlab.io/omni-chat/install-macos.sh | bash
Downloads the latest DMG, copies Omni Chat.app into /Applications,
and runs xattr -dr com.apple.quarantine so Gatekeeper will launch it.
Prefer to install by hand? Open the DMG, drag Omni Chat to Applications, then run
xattr -dr com.apple.quarantine "/Applications/Omni Chat.app"
— or use System Settings → Privacy & Security → Open Anyway.
The one-liner already does this for you.
curl -fsSL https://geekaholic.gitlab.io/omni-chat/install-linux.sh | bash
Downloads the matching AppImage from the GitLab package registry, verifies its SHA-256,
adds a desktop-menu entry, and installs the omni-chat command for your user.
Prefer a portable app? Choose the architecture above, make the downloaded AppImage executable, and run it. Direct download links are refreshed from the release manifest; browse all Linux versions in the GitLab package registry.
# 1. Install Docker Desktop and Ollama first
# 2. Start a small model in Ollama, e.g.:
# ollama pull gemma4:e4b
# Then run oMNI Chat:
git clone https://gitlab.com/geekaholic/omni-chat.git
cd omni-chat
cp .env.example .env
# Tell it to talk to your local Ollama
# (on Docker Desktop this hostname works)
echo 'CHAT_ENDPOINT_URL=http://host.docker.internal:11434' >> .env
docker compose up --pull always -dBarebones chat only. Pulls the image and starts the web app on port 8080.
# After the Starter steps above…
# One-time secret for SearXNG
SECRET=$(openssl rand -hex 32)
echo "SEARXNG_SECRET=$SECRET" >> .env
docker compose \
-f compose.yaml \
-f compose.search.yaml \
up --pull always -dAdds a local SearXNG container so tool-using models can fetch fresh web results.
Power user: full stack
# After clone + cp .env.example .env (and Ollama / endpoint setup)…
# 1) Operator configs (edit after copy — paths, keys, models)
# AI servers and keys: add them after first login in
# Settings → AI Servers (admins) — no file to edit.
cp docs/router.example.json config/router.json
cp docs/mcp.docker.example.json config/mcp.json
cp docs/agents.compose-example.json config/agents.json # Pi agent + SFTP
# 2) Env bits the overlays need
printf 'SEARXNG_SECRET=%s\n' "$(openssl rand -hex 32)" >> .env
# 3) Agent handoff workspace (macOS Docker Desktop example)
docker build -t omni-agent-pi docker/agent-pi
mkdir -p ~/omni-agent-work
# Set managed_root in config/agents.json AND OMNI_AGENT_WORK_DIR to the same absolute path:
echo "OMNI_AGENT_WORK_DIR=$HOME/omni-agent-work" >> .env
ssh-keygen -t ed25519 -f config/sftp_host_key -N ""
# 4) Bring up search + router + fetch + aux + agents + MCP
docker compose \
-f compose.yaml \
-f compose.search.yaml \
-f compose.router.yaml \
-f compose.fetch.yaml \
-f compose.aux.yaml \
-f compose.agents.yaml \
-f compose.mcp.yaml \
up --pull always -d
Layers web search, smart routing, URL fetch, auxiliary model, coding-agent handoff, and MCP sidecars.
Copy examples from docs/ into config/, adapt them, then compose-up.
Needs more RAM/VRAM than Starter — see the
Manual
for memory tiers, Linux DOCKER_GID, and optional imagegen/TTS overlays.
Questions, answered
Anything else is in the online Manual.
Do I need an API key to start?
No. Models served by Ollama, LM Studio, or any OpenAI-compatible server on your computer or home network show up automatically. On a Mac the app can also download a private on-device model, and if you pay for ChatGPT you can sign in with it. Add hosted provider keys only when you want other frontier models.
Can family or teammates share one install?
Yes — that’s what it’s built for. Run it with Docker or on a home server, then add each person under Settings → Members. Everyone signs in from their own browser and gets private chats, projects, and memory, while sharing the models (and any ChatGPT sign-in or API key) you set up once.
How is this different from Open WebUI?
Open WebUI is a great polished Ollama UI. oMNI Chat targets hybrid local + frontier routing, an on-device model in the Mac app, true multi-user accounts, projects/tags, live code previews, inspectable memory, and coding-agent plans in one private workspace.
Will my chats leave my machine?
With local or on-device models, inference stays on your machine or network. Hosted models — including ones you reach by signing in with ChatGPT — only receive prompts when you choose them. Chat history is stored in your local SQLite database either way. Off-the-record mode skips saving entirely.
What hardware do I need?
Any modern computer can run the app. Small open models work around 8 GB RAM; larger models need more VRAM or system memory. On a Mac, the on-device model needs 8 GB, and 16 GB gets the larger 4B version. See the Manual for memory tiers and recommendations.
Is there a desktop app?
Yes. Native apps are available for Apple Silicon Macs (macOS 26 or later) and x86_64 or ARM64 Linux desktops. Install either with the one-line command in Get oMNI Chat. Windows builds are not published yet.
Is there an iPhone or iPad app?
Native iPhone and iPad apps are in testing and coming soon. Until then, if oMNI Chat runs on Docker or a server your phone can reach, open it in Safari and use Share → Add to Home Screen for a full-screen, app-like experience.
Why does macOS warn that the app is unidentified?
The Mac build is ad-hoc signed, so Gatekeeper quarantines a dragged-and-dropped DMG. The one-line installer clears quarantine for you. After a manual install, run xattr -dr com.apple.quarantine "/Applications/Omni Chat.app", or use System Settings → Privacy & Security → Open Anyway.
Where is the full documentation?
The online Manual covers configuration, endpoints, multi-user setup, and advanced features.
Open source. You run it.
Apache 2.0 licensed. Inspect the code, host it yourself, and keep full control of your data and models.