Short answer
SPAWN D is an open-source control plane for command-line coding agents, built by Levy Street. A small Rust daemon on each machine you own dials out to a control server, letting you open shell sessions and run agents such as Claude Code or Codex on that machine from a browser, phone or desktop app. Terminal content travels end-to-end encrypted. It is pre-alpha and invite only.
- Agents run on your hardware under their own login; SPAWN D never holds provider keys.
- The daemon opens no inbound port: it dials out.
- Terminal bytes travel over encrypted WebRTC channels the server cannot read.
- Some metadata, such as working directories and agent definitions, is still server-readable today.
- What
- Open-source control plane for CLI coding agents
- Status
- Pre-alpha, invite only
- Licence
- MIT or Apache-2.0, at your option
- Stack
- Rust daemon, FastAPI control server, Next.js web app, Expo mobile app, Tauri desktop app
Why we built it
AI coding agents work best on a real machine with the real repository, tools and credentials. That machine is often a desktop at the office or a server in a rack, while the person supervising the agent is on a train or in a meeting. Remote access tools were not designed for watching and steering several agents at once, and handing a third-party service your terminal is a bad trade.
SPAWN D is Levy Street’s answer: keep the agents on hardware you control and make them reachable from anywhere, without trusting the middle.
How it works
- Daemon
- A Rust process,
spawnd, runs on each host you pair. It dials out, so there is no inbound port, SSH or VPN to configure. - Sessions
- Real shell sessions that survive daemon restarts, each owned by a worker process with encrypted, bounded replay.
- Agents
- Anything that runs in a terminal: Claude Code, Codex, OpenCode, Aider or a plain shell. Agents use their own logins on your machine.
- Clients
- A web app, an iOS and Android app, and a desktop companion for macOS and Windows.
- Control server
- Handles accounts, pairing, presence and signalling, but not terminal content.
Security, stated precisely
Security claims for developer tools are often vague. SPAWN D’s own trust document names the project’s own servers as the adversary and states what is and is not protected today: terminal bytes, replay, viewport actions and file transfers go browser-to-daemon over authenticated, end-to-end encrypted WebRTC channels, while metadata such as which hosts you own, when they are online, working directories and agent definitions remain visible to the server. That precision is deliberate. It is the same standard we hold client systems to in AI security reviews and audit log design.
Status
SPAWN D is pre-alpha and invite only. Code is dual-licensed under MIT or Apache-2.0 on GitHub. Installation is a single command on macOS and Linux, with Windows supported through WSL.
What it says about how we work
We build our own tools when existing ones do not meet the bar, and we use AI coding agents heavily to do it: much of SPAWN D’s commit history was authored by an agent under engineer supervision. If you are rolling out coding agents in your team, our guide to Claude Code for New Zealand businesses and our notes on agent permissions and approvals are good places to start.
Written by the Levy Street team in Wellington. Last updated 19 October 2026.