Uxnan (pronounced /uʃ.nan/) is a toolkit I'm building to solve a very specific problem I have as a developer: controlling AI coding agents from anywhere, without my hardware becoming the bottleneck.
I work with CLI coding agents (Claude Code, Codex CLI, OpenCode, Gemini CLI, pi-agent) every day. They're extraordinary tools, but the current workflow has real friction:
- When I step away from my PC, I lose all visibility into what the agent is doing. I can't check its progress, approve changes, or send new instructions from my phone.
- Existing desktop solutions are excellent, but many assume high-end hardware. On my current setup, running a heavy IDE + multiple agents + an Electron environment consumes more resources than I can afford.
- There's no provider-agnostic mobile tool that works with any agent, not just one in particular.
Uxnan was born to solve exactly that. It's not an agent — it's the control plane for the agents I already use.
The phone never talks to a cloud middleman in the clear. It pairs with the bridge running on your PC and connects to it directly first — over your LAN or your Tailscale network — and only falls back to an optional, self-hosted relay when you're away from home. Whatever the path, every byte is end-to-end encrypted; the relay only ever sees sealed envelopes.
📱 uxnanmobile 💻 your PC
(Flutter app) ┌──────────────────────────────┐
│ │ bridge ──▶ agent CLIs │
│ E2EE (X25519 + │ (daemon) claude · codex │
├──── Ed25519 + ──────────▶│ opencode · pi │
│ AES-256-GCM) │ gemini │
│ └──────────────────────────────┘
│ ▲
└──── relay (optional, ────────────────┘
self-hosted, off-LAN only —
forwards sealed envelopes, sees nothing)
uxnandesktop — a separate, lightweight desktop app to run & review the same
agents on the PC itself. shared — the contracts both ends speak.
Uxnan is a single repository with five projects. Each one has its own README with the full story; here's the short version and where to go next.
A Flutter app (Android + iOS) that turns your phone into a remote control for the agents on your PC. Watch conversations stream in real time, send instructions, attach images, dictate by voice, review diffs, commit and push, and get a notification the moment an agent finishes — all over the encrypted, bridge-first channel.
A lightweight Agent Development Environment built with Tauri 2, Rust and Svelte 5. Unlike Electron-based alternatives that consume 200-500 MB of RAM just by existing, this ADE uses the OS's native webview and targets 30-100 MB of RAM.
The core idea: each task lives in its own git worktree with its own agent running
in an independent pseudoterminal. I can have 5 agents working in parallel without
one blocking another, switch between them with a click (no git stash, no git checkout), and review each one's changes in an integrated diff viewer
(CodeMirror 6, unified + side-by-side, hunk-level staging) before committing.
It doesn't integrate any agent's SDK. It's terminal-centric: any CLI agent works without modification.
The heart of the product. A small Node.js daemon that runs on your PC, holds the end-to-end-encrypted connection to your phone (E2EE protocol: X25519 + HKDF + Ed25519 + AES-256-GCM), and drives the agents on your behalf by launching each one's official local CLI exactly as you would in a terminal. No provider API, no SDK, no keys: every agent runs under the account you already signed it in with.
Real agents wired: OpenCode, Claude Code, Codex, pi, Gemini CLI.
A tiny, stateless WebSocket relay you can self-host for when your phone and PC aren't on the same network. It forwards sealed E2EE envelopes and nothing else — it never sees your code, your diffs, your keys, or a line of plaintext. Most of the time you won't need it at all.
The single source of truth for the JSON-RPC + E2EE contracts that every part of Uxnan speaks. The bridge and relay consume it directly; the mobile app keeps hand-synced Dart equivalents. If two components ever need to agree on a message shape, they agree here.
- JSON-RPC: envelope types + constructors, error codes (
-32000..-32008+ standard), typed method registry (build-time-locked toMETHOD_NAMES). - E2EE: handshake messages, transcript builder,
SecureEnvelope,PairingPayloadv2 (withhosts: string[]for direct addressing). - Domain models (thread/turn/message, git, workspace, project, auth, session,
approval) and agent contracts (
IAgentAdapter,AgentCapabilities,AgentConfig). - Validation: Ajv-backed validators for requests, responses, E2EE envelopes, pairing payloads and push payloads.
→ Read the shared contracts README
Here privacy isn't a feature, it's the foundation. Everything that travels between the phone and the PC goes through a real end-to-end encrypted channel: session keys come from an X25519 ephemeral exchange, identities are authenticated with Ed25519 signatures, and traffic is sealed with AES-256-GCM. The relay, when it's even involved, is pure transport and never holds a key. Bridge responses are sanitized too — sign-in status, for example, is reported per agent and never returns a token.
If you find a vulnerability, please don't open a public issue — see
SECURITY.md.
-orange?style=for-the-badge){kind=icon}
Uxnan is in alpha, and the core loop already works end to end. From the phone I can pair with the bridge and hold a real, streaming conversation with five agents — OpenCode, Claude Code, Codex, pi and Gemini CLI — over the encrypted channel. The desktop app is alpha-functional on its own. Push notifications are live on Android (iOS depends on Apple assets).
Here's the quick snapshot; the detailed, always-current state for each project
lives in its own FOR-DEV.md:
| Project | Where it stands | Details |
|---|---|---|
uxnanmobile/ |
MVP wired, Android alpha-ready; iOS pending Apple assets | status |
uxnandesktop/ |
Alpha-functional as a standalone app | status |
bridge/ |
Implemented; 5 real agents wired | status |
relay/ |
Implemented; optional / self-hosted | status |
shared/ |
Implemented; contracts locked in CI | README |
This is early software, with no users and no production data yet, so things can still change where a better idea justifies it.
Uxnan is free and open source, built in the open in my own time. If it's useful to you and you'd like to help it keep moving, a coffee goes a long way — and it's genuinely appreciated. 🙏
If you want to build, run, or contribute to Uxnan, everything you need lives off the README so the documentation above stays focused:
CONTRIBUTING.md— how to set up, the quality gates, and how to open a good PR.AGENTS.md— the single source of truth for conventions, architecture rules, and how the docs stay in sync. Read it before any non-trivial change.- Per-project docs — each project keeps task-focused guides in its own
docs/and aCHANGELOG.md:bridge/docs/·relay/docs/·uxnanmobile/docs/·uxnandesktop/docs/. - The specification — the architecture documents are the source of truth for
cross-component behavior:
architecture/(mobile, bridge, relay, shared) anduxnandesktop/architecture/(desktop). - Releases & versioning —
VERSIONS.md.
Uxnan is released under the Mozilla Public License 2.0.
Uxnan — a name with no relation to, or derivation from, any existing product.
