-
Notifications
You must be signed in to change notification settings - Fork 0
Home
In short. AgentBox gives every AI agent on your machine one place to reach you: a card over whatever you are doing, a sound that says what kind of thing arrived before you look, and an answer that goes straight back to the code waiting on it. It exists so you can leave an agent running for an hour and still be the one who decides what matters.
Read on if you are not yet convinced an agent can be left alone. Skip to Safe on a work machine?, or Install.
An agent works for twenty minutes while you are somewhere else. Then it needs one word from you.
It has nowhere to put the question, so it does one of two things, and both cost you. It stops and waits in a terminal nobody is looking at. Or it decides for itself, and you find out afterwards, in the diff.
That is the problem. Not the model, not the prompt. The last meter.

Every option carries its number. The whole exchange is one keystroke.
Every option carries its number. The whole exchange is one keystroke.
The card appears dead centre of the monitor your pointer is on, above every window, and it does not take your keyboard. That last part is the decision the rest of the product is built around: a card that grabs focus while you are typing can swallow a password, or let a stray keystroke approve a production deploy. So your typing keeps going where it was going, and when you are ready you press the summon key and answer with one digit.
Answers wait three seconds behind an undo strip before they are sent. A mis-click costs nothing.
And when a second agent asks something while that card is up, the footer says so. You are never told about one question and quietly left with two.
Most of what an agent has to say is not a question, and interrupting you for all of it would train you to ignore the one that matters.
So there are five levels and six short sounds, the sixth because a question gets its own chime whatever level it carries. Info and success take themselves off the screen after six seconds. Warnings, errors and anything urgent wait to be read, because a notice with no deadline was worth sending. Urgent is the only level that pierces do not disturb, and you can switch that off too.
An agent may also attach one spoken line to anything it puts on screen, read out after the chime. The sound tells you roughly how much something matters. The sentence tells you what it was, without your eyes leaving what you were doing.
Everything that arrived while you were gone is in the inbox. Still-pending questions first, answerable right there with the same keys a card takes, then the day's history: who asked, what for, and how long you took to answer.
Nothing truncates there, and that rule was bought the hard way. A countdown card with a 946-character body expired unanswered while the desk was empty, and all that survived was 140 characters in a tooltip. For a tool whose whole purpose is that a message is not lost, losing the message is the wrong failure. Every row now opens and reads its item back whole.
With several agents running there is a board that says which of them is asking you something, which is stuck behind another one, which is deliberately waiting, and which has gone quiet long enough to worry about.
Three questions, and straight answers.
Can it leak a secret? A credential goes from your keyboard into a 0600 file
and the agent is handed a path, never the value. The log records that a secret was
asked for. It has never recorded one.
Can agent-authored content phone home? No. An interface an agent writes runs with no network at all, and an image in agent prose may name a file on your disk and nothing else. Both are enforced in the code that renders them, not by a promise on a page.
Can I switch it off? Do not disturb, per-agent mute, quiet hours, and a gate that holds everything while you are away or presenting. Urgent is the one level that still gets through, and even that has a switch.
Important
AgentBox builds and runs on Linux, macOS and Windows, and nothing that carries a message needs a display server. X11 is where it looks its best: it is the one desktop where AgentBox places its own windows, so a card lands dead centre of the screen you are on and appears above without taking your keystrokes. Elsewhere the window manager places it and it takes focus. Two capabilities are X11-only and say so when asked - the global hotkeys and driving the desktop. Nobody has run the macOS build yet. The whole picture.
One Go binary and a resident daemon that owns the windows. Thirty-nine tools over MCP, which is how a coding agent reaches it, and a shell command for nearly all of the same jobs, which is how hooks, Makefiles and cron reach the same daemon. A unix socket with a peer check on every connection, no network listener of any kind, no account, no telemetry, nothing leaving the machine.
Because a blocking question is also a shell command, a human answer is an exit code:
if agentbox confirm --title "Push to main?" --body "12 commits, tests green"; then
git push
fiNo parsing. 0 is yes, 1 is no, 3 is nobody answered.
Next: what a question actually looks like, or straight to Install.
AgentBox
Decide in five minutes
Living with it
Several agents at once
AgentBox runs agents too
Look it up
The agent-facing reference is not here. It ships inside the binary:
agentbox docs agent.