Trying it

There is no hosted trial. You install Bee Box on a machine you control, then create a box: your own directory, kept under version control with git, holding the cards the system works with.

The container path is the primary one. Docker with Compose v2 plus a clone of the repository, because the guide builds the image from that clone (a prebuilt image is published daily to GitHub's container registry, and the guide does not yet cover pulling it). The sequence is four steps: create the box in a directory you choose, sign in to the coding agent, start the container, then open the printed local URL (the guide gives the exact commands). By default the box listens only on your own machine, so nothing else on your network can reach it until you decide otherwise. Details: install/docker.md. That guide documents the Claude login and the Codex login. Accounts, stored provider keys, and the Codex login live in a second named volume, so a rebuild keeps them; the first start also downloads about 2 GB of document-reading models in the background.

The from-source path is for modifying the engine. It needs Node 24, pnpm, and several system binaries, and it ends with a doctor command that checks every prerequisite and prints a remedy for anything missing. Details: dev/developer-install.md.

The agent-driven install. The project publishes an install guide written for your own coding agent, so you can ask Claude Code or Codex to do the install with you: install/agent-install.md. That guide tells the agent to treat the instructions as advice rather than authority, and to ask you before anything that affects your machine, your data, or your accounts. Four decisions are yours: run it or hack on it, where the box lives, your own agent login, and whether the box is exposed beyond loopback.

The first hour. A new box opens on a chat with one question beside it, "What do you want to keep track of here?" Answering it in your own words is how the box learns what it is for; "Later" puts it away, and the question stays answerable until you answer it. Then start with the inputs that need no external account: type into chat, record a voice memo, photograph things and talk about them. Scheduled syncing of email, calendar, and Drive is off on a fresh box, so no outside content arrives until you turn it on. A few housekeeping routines are on, and the three that run an agent (weekly retrospective, daily todo review, daily question review) start one only when a quiet chat, an overdue todo, or a waiting question gives it something to do. Connecting Google, which unlocks Gmail and Calendar, is the slower step; the documentation suggests leaving it for later.

Updating means fetching the latest version of the project and rebuilding. On start, the container migrates the box's card data and refreshes its generated agent docs. See status and maturity, which also notes what the update story does not yet cover.