Shipping

GaleChat · streaming replies, vision, open models. CirrusCode · inline diffs, repo-aware edits, PRs. AsheAgent runtime · scheduled hands, activity feed.

Beta & upcoming

MistSearch · local + web fusion, cited answers. StratusVision · OCR and capture-to-ask. OvertureAudio · transcription, dictation, voice-first. HazeMemory · private vector recall, per-chat. ZephyrTranslation · on-device multilingual drafts.

Weather, by design

8 modules, one engine.

Every module runs on the same on-device runtime — swap models, share context, and nothing leaves your phone.

Runtime

On-device engines MLX, GGUF via llama.cpp, Core ML — selected per device. Capability manifest Model routing by vision, PDF, OCR, tool-use flags. Device fit Auto-tiered by chip, RAM, and measured performance.

Experience

Unified theming One palette across modules — vanilla wood by default. Privacy posture Local-first, zero telemetry, opt-in network. Open models Browse Hugging Face, install, swap, uninstall.

Local-first

Nothing leaves your device.

Every inference, every file, every log — all of it stays on-disk. No account, no sync.

Learn

Blog Notes from the build — runtime deep-dives, model picks. Changelog Every shipped module, model, and fix. Roadmap What's queued, what's shipping next.

Build

Docs Runtime concepts and module primers. Model catalog Curated, device-verified picks for each module. Community Questions, tips, and model recs from power users.

From the blog

Picking a model that actually runs on your iPhone.

A look at the capability manifest, memory budgets, and why a quantized 3B beats a streamed 70B every time.

About

Who we are A small team building quiet, local-first software. Mission AI that you own — not rented by the token. Press Brand, logos, screenshots.

Policies

Privacy What's collected (nothing that leaves your device). Terms The short, human-readable agreement. Security Sandbox, keychain, and network posture.

On device

Private by construction.

No accounts. No keys. No telemetry. The network off-switch is the default.

on-device · private · local-first

On-device intelligence.
Chat, code, and agents — all running locally, nothing leaves your iPhone.

Coming to App Store See it in motion

iOS 17+iPhone & iPad

100%on-device inference

Open modelsMLX · GGUF · Core ML

Meet Nimbus8

A private AI studio that lives on your device — not in a datacenter.

Every model runs locally. Every chat stays on your phone. Built from scratch in native Swift with a Rust core, Nimbus8 is the quiet, careful place to talk to open models without a subscription, a login, or a round trip to the cloud.

Gale · Chat

Gemini‑class chat. Running on your iPhone.

Gale is the everyday assistant. Pick any open model — Llama, Qwen, Gemma, Mistral — and talk to it with the polish of a modern chat app. Streaming thoughts, attachments, tool calls, all local.

Streaming repliestoken-by-token, with live "thinking" blocks you can fold away
Photos · PDFs · filesvision and OCR routed by a capability manifest per model
Zero cloudno accounts, no keys, no telemetry

9:41

Nimbus8

Good afternoon, Cameron

Where should we start?

Plan me three realistic priorities for this morning, plus a playlist.

Thinking

planning your morning

Ask Nimbus8

9:41

Cirrus

anthropics/claude-cookbook

base: main

›

Find the rate-limit bug in requests.py

Read requests.py

128 lines scanned

Found it — the exponential backoff resets on every retry instead of compounding. Here's a patch:

requests.py +4 −2

@@ -42,6 +42,8 @@ def fetch_with_retry(url):

for attempt in range(max_retries):
    try:
        return http.get(url)
−   except RateLimit:
−       time.sleep(1)
+   except RateLimit:
+       delay = base * (2 ** attempt)
+       time.sleep(delay + jitter())
+       continue

Message Cirrus…

Cirrus · Code

Review code like a PR. Ship it like a dev.

Cirrus plugs straight into your GitHub repos. Describe a task, get an inline diff card with line numbers, hunk headers, and a one-tap "Open PR". All the edits are produced on-device — your source never leaves the phone.

Inline diff cardsGitHub-style gutter, +/- counts, expand/collapse
Repo-awarepick any branch, scoped to files you approve
Editor tabread-only inspector with syntax highlighting

Ashe · Agent runtime

Hands that do things. On their own. On your terms.

Ashe is Nimbus8's on-device agent runtime. It runs scheduled "Hands" — little workflows that call tools and models on a tick, then show you exactly what they did in a live activity feed. Every destructive action asks first. Nothing runs in the cloud.

Activity feedlive stream of every tool call, model call, and tick
Handsscheduled tasks you enable one at a time, with manual overrides
Confirm-firstwrites, network, disk — always opt-in, every run

9:41

Ashe

Activity

Live tool calls, model calls, and hand ticks.

Invoked skill · inbox.triage

3 emails classified

Ran hand · morning-brief

2.4s · model: Llama-3.2-3B

Thinking

drafting summary…

Needs confirmation

write to ~/Nimbus8/brief.md

Hands

3 enabled · next tick in 6m

morning-brief

every 1h · last: 2m ago · 0 errors

inbox.triage

every 15m · last: 14m ago · 0 errors

Waiting for Apple.

Nimbus8 is wrapping up with Apple. Drop your email and we'll send you an early-access invite the moment the build goes live — no waitlist theatre, no marketing list, just the link.