Sail

sail

Sail turns bare-metal servers into isolated, spec-driven dev environments where AI coding agents do the work. Each project gets a hard-isolated container. The team's specs and project definitions live in a synced database. Agents are dispatched against those specs with rollback safety and cross-agent review.

Sail is not a coding agent. It is the layer around them: the environments they run in, the specs they pick up, the reviews they pass, and the shared state a team works from. It does for coding agents roughly what Kubernetes does for containers, orchestrating the work rather than performing it.

Built with Java 25, picocli, and GraalVM native-image. One binary, no runtime dependencies, sub-millisecond startup.

Install

curl -fsSL https://raw.githubusercontent.com/standardapplied/sail/main/install.sh | bash

`sail upgrade` replaces the binary in place and converges the database. Linux (amd64) runs a full host. macOS (arm64) runs as a thin client that drives a remote host over SSH.

Upgrade compatibility

Version 0.15.0 is the v1 upgrade floor. Upgrade every main and node box to 0.15.0 before installing any later v1 release. A mixed fleet below that floor is refused before sync touches data, with an error naming the required version.

The model: one main, many nodes

An org runs one main box. It holds the team's specs and project definitions in a SQLite control plane and is the source of truth. Every engineer also has their own box. A box pointed at main is a node: it pulls specs, project definitions, and shared files from main, and pushes its own work back.

There is no GitHub in this loop and no separate board. The database is the board, and `sail sync` moves it over a locked-down SSH gateway using public-key auth, with three-way conflict resolution so no one's work is overwritten. Compute is never scheduled across boxes. The star coordinates state, not execution.

Quick start

Stand up the main. One idempotent command provision