Zenith runs a full AI workspace — chat, a coding agent, deep research, long-term memory, image generation, voice — entirely on your own machine, against your own Ollama models. Nothing you type is ever sent anywhere.
Why Zenith
Every message, every file, every question — sent to a server you can't see, kept under terms you didn't write. Zenith doesn't work that way.
It runs entirely on hardware you control, talking only to Ollama — a free, local model runner — on the same machine, or your own network. No account. No subscription. No API key, ever.
Feature tour
Not a chatbot demo — a daily-driver, verified against a real running deployment.
Streaming chat, images, documents, video — routed automatically to whichever installed model fits the message.
Shell, file edits with diff previews, persistent sessions, real git — with an approval gate from manual to full-auto, your call.
Multi-step web research that plans, searches, and writes a cited report — refuses to answer before it's actually searched.
Learns durable facts about you and recalls them automatically — with a conflict review for things that no longer add up.
A real file tree, editor, and git status scoped to a working directory — extract a conversation's code into a runnable project in one click.
See how conversations, memories, and documents actually connect — built from real relationships, not a guess.
Trigger → action rules on real backend events, plus outbound webhooks so your own scripts can react to what Zenith is doing.
Hold the mic to talk, hear replies spoken back — including a hands-free continuous conversation mode.
Bring your own local Stable Diffusion server — generated images render right in the chat.
Connect external Model Context Protocol servers instead of waiting for a built-in integration — a curated starter list included.
Real API-level lock middleware for shared machines, plus AES-256-GCM encrypted credentials at rest.
A real installable app with a system tray, global hotkey, and auto-update — not just a browser tab.
Who it's for
Zenith has a lot of surface area (see the feature tour above). Nobody uses all of it. Find yourself below and turn on just that subset.
You want everyday conversation — typed or spoken — that isn't training a model somewhere else. Nothing fancier than that.
Skip: agent mode, MCP, the code editor — none of that shows up unless you go looking for it.
You want a coding agent with real teeth — one that can touch your filesystem and git history, not just paste snippets into a chat box.
You're working from sources — papers, drafts, a growing pile of context — and need the app to hold that across sessions, not just one chat.
You picked Zenith specifically because nothing leaves your machine by default — and you want to know exactly where the few exceptions are before you flip anything on.
You want the whole instrument panel — rules that fire on real events, tools wired in from elsewhere, and access from more than one device.
Get started
Pick whichever matches how comfortable you are with a terminal.
docker compose up -d --build and you're done.msi / Linux .AppImage, built by CIBefore you install
Zenith itself is lightweight — the real requirement is whatever Ollama model you plan to run. Settings → Model routing scores your actual installed models against your real GPU/RAM automatically; this is the general shape of it.
| Your GPU VRAM (or RAM, CPU-only) | Comfortable model size | Still usable |
|---|---|---|
| No dedicated GPU / under 4 GB | 1–2B params | up to ~3B |
| 4–8 GB (e.g. RTX 3050/3060 laptop) | 3–4B params | up to ~8B |
| 8–16 GB (e.g. RTX 4060 Ti/4070) | 8–12B params | up to ~20B |
| 16 GB+ (e.g. RTX 4090, workstation) | 20B+ params | whatever fits |
Rough heuristic (~0.65GB per billion params at Q4 quantization, plus runtime overhead) — not a promise. A model too big for your GPU doesn't error, it just spills onto much slower CPU inference; Zenith's own stuck-generation watchdog catches and reports this instead of leaving you staring at a spinner.
FAQ
Yes — Zenith is open source under AGPL-3.0, no subscription, no paid tier, no usage limits. The only ongoing cost is your own computer's electricity. Optional features that touch the open web (search, page reading) use free, keyless endpoints, not paid APIs.
No, by default. Zenith only talks to your own local Ollama instance and, only when a feature that needs it is explicitly turned on for that message, the open web (search results, a page you asked it to read). There's no telemetry, no analytics, no account.
No — it runs on CPU-only setups too, just slower. See the hardware guide above; even modest laptop GPUs run small models comfortably.
Yes — it's an installable PWA with a mobile-friendly layout. Open it from your phone's browser over your LAN or Tailscale and add it to your home screen.
AGPL-3.0-or-later: use, modify, and self-host it freely. If you run a modified version as a network service others can use, you have to publish your changes too — a deliberate choice, matching the project's "no cloud middleman" philosophy.
No — Zenith is an independent, unaffiliated client that talks to Ollama's own local API, the same way any other app on your machine would.