SETUN LABS
// APPS ◈ POCKET AI HUB

AI that lives
where your data does.

A phone app plus a PC companion, DESK console, CLI, and daemon, that run AI entirely on your own hardware: on the phone itself, on your PC over your LAN, or both at once. No accounts, no cloud required, no telemetry. Measured, never estimated.

⬢ STATUS // Feature-complete and in final on-device testing. No public download yet, the lists at the bottom get it first.

0 FEATURES
COUNTED, TESTED, AND DOCUMENTED IN THE FULL FEATURE AUDIT
// THE PHONE APP
// 01

Chat

チャット

Streaming conversations with local models, with the full toolbox.

  • History with search, regenerate, edit-and-resend
  • Markdown, image attachments (up to 4 per turn), export
  • Per chat, pick the source: this device or your PC
  • JSON mode, personas, per-model sampler configs
// 02

Models

モデル

Your phone is profiled and the catalog is ranked by what actually fits it.

  • Device profile: RAM, chip, storage, acceleration
  • Recommender ranks live Hugging Face models that fit
  • Real measured speeds after you run a benchmark
  • Resumable downloads with honest progress; sideload GGUFs
  • Tool-competence probe: can this model drive tools, on your device, not someone's chart
// 03

Pair

ペアリング

Link phone and PC over your LAN in one code.

  • mDNS/UDP discovery or manual entry
  • 6-digit verified pairing code
  • TLS with fingerprint pinning + cert-change detection
  • Plaintext consent gates, quick-pair resume
  • One tap to unpair or revoke
// 04

Devices

デバイス

The device-pool dashboard.

  • Every linked device: name, live status, battery, RAM, link health, measured speed
  • Honest fit / marginal / no verdict on joining work, with the math shown
  • Open, documented read-only API: hubctl devices metrics --json
// 05

Agents

エージェント

Three modes.

  • A labeled cascade demo
  • Real supervision of your PC's coding agents: watch live, approve/deny with countdowns, steer, start per project
  • Task scheduler: queue prompts for later, results pushed to your phone
// 06

Tools & Safety

道具と安全

The AI can act, under your rules.

  • Actions: shell, files, launch apps, web search, reminders, your own custom tools
  • Four-tier permissions: yolo / ask / deny / custom
  • Approval cards on your phone
  • Append-only audit log, viewable and exportable; failures in plain language
// 07

Memory & Files

記憶とファイル

Your context stays yours.

  • Opt-in cross-chat memory: you pin what's remembered, wipe anytime
  • On-device RAG: point the AI at your folders and answers quote your own documents, sources shown
// 08

Settings

設定

Yours to shape.

  • 10 theme presets + a custom editor
  • Reduced motion, power modes, battery-exemption guidance
  • Consent-gated phone control, notifications, tutorial replay
  • Catalog management, benchmark export
// THE DESK CONSOLE ◈ PC
THE CONSOLE A full PC console: hub start/stop, live pairing codes, connected phones, relay activity, AI source management (auto-detect + auto-adopt, OpenAI-compatible, opencode, Claude Code, any CLI, custom tools), tool-security editor, supervision deck (watch / approve / steer), log tail, audit deck, phone-chats reader, docs reader, themes, and the same tutorial.
HUBCTL Every one of those functions, scriptable from the terminal.
// THE HARD PARTS NOBODY ELSE SHIPS
PHONE AS AI SOURCE Opt in, and your PC runs prompts on your phone's model over the link, one engine, shared honestly.
DEVICE POOLING Tensor-parallel inference across your phone + PC via llama.cpp RPC, with honest measured numbers, and honest "not worth it" verdicts when it isn't.
AGENTS CONTROLLING PHONES Consent-gated screen read plus tap / type / swipe from a PC agent, audited.
CASCADE JUDGING A small model answers; a calibrator judges confidence; weak answers escalate to a bigger AI, with an honest UNRATED when judging can't be proven.
ENCRYPTED, AUDITED, HONEST TLS with pinned fingerprints, permission tiers, decision TTLs, an audit ring, and a UI that says what's true, measured, never estimated.
// OPEN-CORE ◈ STATED PLAINLY

Everything you run is open, FSL-1.1, converting to Apache-2.0 two years after each release. Here is exactly where the line falls:

Open · the app The entire mobile app (Android; iOS in progress): chat, models and recommender, downloads, benchmarks, pairing, device dashboard, agents, settings, themes, tutorial, RAG, vision, memory, the phone-control stack. Every screen and every animation is in the repo. ◈ FSL-1.1
Open · the desktop The DESK console (all nine decks), hubctl (every command), and hubd, the daemon itself, the heart of the system. ◈ FSL-1.1
Open · the protocol Every wire frame is documented, pairing, chat relay, tools, decisions, tasks, audit, sync, device metrics, phone generation, phone control, the pool. Anyone can build a client against it. ◈ DOCUMENTED
Open · data & seams The device-metrics API is stable, documented, read-only. The benchmark JSON schema, the CascadeJudge interface, the tool-pack format, the CLI registry, the catalog file, all open and user-editable. ◈ YOURS
Closed · two things Deliberately: the calibration tuning behind CascadeJudge (the interface is open; the tuning is the trade secret), and the future hosted benchmark-aggregation service (nothing in the app depends on it; no open feature is gated on it). ◈ THE BUSINESS
The rule If a closed piece vanished tomorrow, the open product loses nothing. ◈ HELD TO

⬢ FINE PRINT // FSL-1.1 means source-available now, Apache-2.0 automatically two years after each release, you're never locked in forever. "Open" covers code and protocol: your API keys, models, chats, and device names never touch anything we run, there is no server. The mobile and desktop apps are one codebase sharing one design system, so "open" means identically open on both.

STAY CLOSE近づいて

Hear it from us when it launches.

PLATFORM // ANDROID FIRST STATUS // FINAL ON-DEVICE TESTING COST // FREE LISTS, NO SPAM