What is Agent Island?
An open-source macOS utility for Claude Code and Codex users. It watches long-running local agent sessions and makes finished, stalled, rate-limited, or waiting states visible.
Open source / MIT / v1.2.1
Agent Island
A local night-watch for long Claude Code and Codex runs.
It watches the session state, keeps quiet while work is moving, and makes the handoff visible when the agent is done, stalled, rate-limited, or waiting for you.
Repository proof
An open project first, a product site second.
The important claims point back to inspectable public surfaces: source, version tag, license, docs, issues, and collection listings.
github.com/tristan666666/agent-island
Open repo
source build
git clone https://github.com/tristan666666/agent-island.git
cd agent-island
./scripts/verify.shHandoff loop
The page follows the same state change as the product.
Scroll or click the states. The copy and progress rail move from quiet monitoring to handoff, warning, and trusted resume.
01
Working quietly
The transcript is still changing, so the provider logo rotates in the notch instead of interrupting you.
Product surfaces
Real app surfaces, framed as evidence.
Trust boundary
Local-first, with the sharp edges named.
- Local Mac utility.It runs on your machine, not as a hosted session service.
- Session-state reading.It reads local transcript and usage signals to detect working, done, stalled, or auth-required states.
- Explicit resume risk.Auto-resume uses dangerous permission-bypass flags and belongs only on workflows you already trust.
- Inspectable update path.Sparkle notes, version tags, license, and security policy are all public in the repo.
FAQ
Answers before you install.
The short version of the project docs: what it is, where it runs, what changed in v1.2.1, and what needs trust.
Does it run in the cloud?
No. It is a local Mac app. Session state and trigger run logs stay on your machine unless you decide to share them.
What changed in v1.2.1?
The v1.2.1 version centers on live session states, a foreground turn alarm, auto-resume controls, and a status guide for the attention states.
How does detection work?
Agent Island reads local transcript and usage signals conservatively, then maps them into working, your-turn, stalled, rate-limited, or auth-required states.
What needs extra trust?
Auto-resume uses permission-bypass flags. Attach triggers only to repeatable local workflows you already trust.
How are updates and security handled?
The version tag, Sparkle update notes, license, and security policy are all published in GitHub so they can be inspected directly.