singular / sing
Open Source

Isolated
environments
for AI agents.

A single native binary that provisions bare-metal servers and creates fully isolated development environments.

View on GitHub MIT License
sing
1
Binary
0
Dependencies
90s
To ready
// the problem

AI agents need to act without asking.

They install packages, run builds, execute tests, and modify code autonomously. When you're running multiple projects with agents grinding overnight, you need more than Docker namespaces.

You need full system-level separation where one runaway agent cannot touch another project's code, database, or runtime. That's what sing provides.

// architecture

System containers. Not application containers.

Each project gets a full Linux userspace with its own filesystem, network stack, and systemd — isolated by cgroups and managed by Incus.

BARE-METAL HOST
project-alpha
jdk 25 node 22
postgres:16
redis:7
meilisearch
claude-code
project-bravo
rust python 3.12
mysql:8
rabbitmq
autonomous overnight
sing project create ...
Incus

System containers with full Linux userspace. Each one looks and feels like a dedicated machine.

Rootless Podman

Daemonless containers inside each sandbox. No privileged root process. Services invisible across projects.

Cgroup Isolation

Hard-capped CPU and memory per project. One agent burning cores doesn't affect the others.

// capabilities

Everything an agent needs.
Nothing it shouldn't have.

01

Snapshot & Rollback

Every autonomous run begins with a container snapshot. If an agent goes sideways overnight, you don't debug a corrupted environment — you roll back to the exact state before it started. Recovery is always ninety seconds, not ninety minutes.

02

Guardrails

Wall-clock time limits. Idle detection. Commit frequency monitoring. All enforced from the host, outside the agent's container — outside its reach. If something spirals, the watcher stops the session and rolls back automatically.

03

Spec-Driven

Work organizes around version-controlled specs with dependency tracking and assignee enforcement. Write the spec during the day. The agent executes it overnight. Review the results in the morning. A repeatable loop.

04

Context Generation

Generates CLAUDE.md, AGENTS.md, SECURITY.md from your config. Agent-agnostic context files that commit alongside your code and persist institutional knowledge across sessions.

05

Multi-Agent

Configure different agents for implementation and review. Claude Code builds, Codex reviews — or the other way around. Reviews trigger automatically when a spec completes. Cross-agent orchestration is built in.

06

Zero Dependencies

Built with Java 25 and compiled to a native binary with GraalVM. Sub-millisecond startup. Installed with curl. No runtime to manage. No containers to run the tool that manages your containers.

// get started

Install in one line.

$ curl -fsSL https://raw.githubusercontent.com/singlr-ai/sing/main/install.sh | bash