Five Claude Code sessions, two Codex runs, one brain. Silicon Workers lives in your menu bar and watches every session for you — so you know in 3 seconds: who's waiting, which one is urgent, and what to do next.
Free · Open source · Your data never leaves your Mac
AI coding agents turned everyone into a multi-threaded manager. But the tooling is still single-threaded — you're managing a fleet with Cmd-Tab and memory.
Did it finish yet?You switch over — it finished 20 minutes ago and has been idling, waiting for you.
Which window wanted approval?Three terminals, two desktop apps. You check them one by one to find it.
Which one do I look at first?Everything looks unread, but only one is actually blocked on your decision.
The top of the panel holds exactly one focus card: the task that needs you most. Approvals first, then whoever has waited longest. Open the panel — the first thing you see is the answer.
"Needs your approval" means the agent is truly stuck — solid orange, counted on the menu bar badge. "Finished, take a look" can wait — soft orange, no badge. Every number on the badge deserves your hands right now.
Claude Code reports through official hooks, every event. Codex reports each turn via notify. Anything else, register by hand. Each project gets its own color — mixed fleets stay legible.
Three tiny workers appear in your menu bar. No Dock icon, no Cmd-Tab entry — quiet enough to forget.
Click "Connect" at the top of the panel — Claude Code and Codex event capture installs itself. It only writes a log, never changes behavior, and uninstalls cleanly.
Open that fifth session with a clear conscience. When someone's waiting on you, the badge counts it — click and you know who's first.
No server, no account, no telemetry. Event capture appends one line of JSON to a local file; Silicon Workers assembles the state on your machine. Uninstalling means deleting two lines of config.
No. Silicon Workers watches, never touches — its hooks produce zero output to Claude Code, so approval prompts reach you exactly as before. Its only job is telling you one is waiting.
Close, but honestly no. There's no universal "the AI is waiting for input" event — only approval requests are precise. Killed terminals are caught by timeouts. Our principle: better to nudge you once too often than pretend to be omniscient.
Not yet. V1 covers Claude Code (all events) and Codex (turn completion); anything else can be registered manually. Want your agent on the roster? Open an issue on GitHub.
Free and open source during beta. The core will stay free.