Why not just use a chatbot?

Bee Box rents its agent loop from Claude Code or Codex, so it runs on the same models the chat apps sell. It is a wrapper, and anything below could be added to a chat app. The question is what Bee Box does today that they do not, as they are generally offered.

A box is one directory of your data, kept under version control with git. A card is a file in it with a structured header and, usually, some text. The agent is the coding agent that operates the box (glossary).

Against a chat app

• Memory is files you own. What it remembers is cards on your disk, and every change is a git commit you can read and trace (design/durability-and-provenance.md).
• It acts between conversations. It is idle by default and wakes when a schedule fires or a connector pulls new data. You can leave a voice memo and walk away (capabilities/schedules.md).
• Connectors make records. Email threads, calendar events, and documents sync in as typed cards it can re-read and act on months later, rather than text in one conversation.
• Corrections become rules. Fix a filing decision or answer a question, and the correction is written where the agent reads it next time (design/teaching.md).
• It keeps your words and their origin. What you said is stored as a quote, and a fact drawn from an email or a page points back at it, so the agent's summary and your own words never blur together. See provenance.
• It stays coherent as it grows. Links between cards are parsed and checked, references are rewritten when a card moves, and every card is validated at several points before a change is kept, so years of accumulation stay a navigable hypertext rather than a pile of files (integrity).
• It asks instead of guessing. Lacking confidence or authority, the agent writes a question card and waits for your answer (capabilities/questions.md).
• It builds inside the box. It has shell and filesystem access on your machine, so views, small programs, and new card types it makes stay there.
• Capture lands in the same store. Photos, voice, and clipped web pages from a phone or browser become cards in the one box (capabilities/phone-capture.md, capabilities/web-clipping.md).
• Nothing is hosted. The data stays on your machine, and so does the risk (10-your-data-and-safety.md).

Against Claude Code or Codex on their own

If you already keep instructions and scripts for a coding agent, some of this you could assemble yourself. What Bee Box supplies is mechanism rather than instruction: each kind of card has fields that get checked when it's created, and checked again automatically before every change is saved, so a malformed record can't slip through; the intake and triage pipeline; the question loop; connectors; the scheduler and procedures; a web interface with chat; phone capture; and a reference corpus generated from the engine source and handed to every box agent (reference/index.md). Instructions still carry a lot, so this is a difference of degree.

What the chat apps do better

They are polished, have native mobile apps, and take no setup. Their memory works the moment you sign in. They have voice modes, a wide range of integrations, no machine to keep running, and no failure that is yours to debug. Bee Box asks for a computer that stays on, an agent subscription, and a tolerance for files and git. If what you want is a chat assistant that remembers you, the app is the right answer.

What people use it for is in 02-what-you-can-use-it-for.md; other agent systems are compared in 12-compared-to-alternatives.md.