← What an agent may and may not do
What can my AI see and do in my vault, and does it remember?
Short answer
In GilbApp's chat your AI works as you; in Terminal, only in vaults granted to it. A chat ends with its tab.
Before you start
Words used here (1)
- GilbApp
- GilbApp is the app, opened in a web browser at https://gilb.com/app, where you open, fill and share your vaults.
Deeper
Also asked:
Does my AI see all my data? In GilbApp's chat it works as you, so it can reach whatever you can reach. In Terminal it reaches only its own AI vault, the vaults it made there and the vaults you gave its AI vault access to.
Can my AI read my other vaults without asking me? In Terminal, no: it reaches only the vaults you granted to its AI vault, and those it made. In GilbApp's chat it works as you, in the vaults the screen lists unless you name another, so it can reach any vault you can.
Your AI reaches your vaults in two ways, and they differ.
In GilbApp's chat, the agent you choose works in your vaults as you, with your own access, in the vaults the middle of the screen lists, unless you name another vault in your message. You can ask it to list your vaults, find and read records, and create records and the links between them, and delete records; it can reach only what you can reach. It does not reach your computer: with Claude Code, the bridge, a small program on your computer that links the chat to your Claude Code, runs Claude Code with its own tools off (no shell, no files, no web), so it does not reach the files on your Mac, unless you choose a code folder and approve: then it works in that folder only. The answer How do I let my AI build in a folder on my computer from GilbApp's chat? shows how.
In Terminal, the Mac's window for typing commands, Claude Code works through its AI vault, a small vault that stands for your AI. Your AI reaches its own AI vault, any vault it makes there, and the vaults you give its AI vault access to; with no grant it reaches none of your vaults. read lets your AI read what is in the vault. read gives no writing, so for it to add or change records you give write too. There, Claude Code can also work with files on your computer. Anthropic says that in Manual mode it "starts with read-only permissions. When it needs to edit files, run tests, or execute commands, it asks you first", and that in Auto mode "A separate classifier model reviews actions instead of you and blocks the ones it judges unsafe" (https://code.claude.com/docs/en/security, read 13 November 2026).
In either way, the tools GilbPlatform gives your AI cannot delete a vault, and cannot give or take away anyone's access. From Claude Code it can make a new vault only while This AI may create vaults is ticked; in GilbApp's chat it makes one as you when you ask for one.
What it remembers: the conversation in GilbApp's chat stays in this browser tab only, and only its last 32 messages: closing the tab ends it. A new chat does not know the old one. What lasts is what is saved in your vaults. Kai-Zen keeps each project's files, its specs, reports, findings and handoffs, in the project's own folder; a handoff is the note Kai-Zen writes when you end a session: what was done, what is open and where to start next time. You do not need your earlier chat: Kai-Zen reads the project itself. In Claude Code, a new chat starts with no vault chosen, so name the vault you want to work in again; only when your AI can reach just one vault besides its own is that vault chosen by itself. Claude Code itself starts each session with a fresh context window; Anthropic says two things carry knowledge across sessions: CLAUDE.md files, which you write, and auto memory, notes Claude writes itself (https://code.claude.com/docs/en/memory, read 13 November 2026). Those are features of Claude Code.
What your AI reads from your vaults also goes to the company that makes it; the answer on whether your AI company sees your data says more.
This answer is written for a computer screen.
Source: GilbPlatform 670005cc (https://id.gilb.com/version read 12 November 2026): src/ai/vault-tools.ts:907-1013 (the AI's vault tools: list_vaults, list_projects, list_records, read_record, list_types, create_record, list_predicates, create_relation, read/write_plan_card, read/write_plan_table, read/write_project_outcome, delete_record; no tool deletes a vault or changes access), :719-723 (delete_record ends a record a person can write; history remains), src/ai/chat-vault-tool.ts:25-26 (create_vault, check_vault_name). Connector gilb-connector 0.5.15 README (vendored in GilbApp 96968de7): 0.5.15 note (a new chat starts with no vault); "If this AI can reach exactly one vault besides its own, that vault is selected automatically."; It cannot create a vault unless the person allows it. Sentences reused word for word from the live answers How do I chat with an agent in GilbApp? (works as you; 32 messages), How do I connect my AI to GilbApp chat? (bridge tools off, unless you choose a code folder), How do I give my AI access to a vault? (reach; read; read gives no writing), What is an AI vault? (This AI may create vaults), How do I continue a project with Kai-Zen? (project files; handoff; earlier chat). Anthropic: https://code.claude.com/docs/en/security, read 13 November 2026. Code-folder exception: GilbApp 49b1ae53 (https://gilb.com/app/version.json, read 13 November 2026), src/components/agent/AgentChat.tsx:677-692 and :1084-1090 (Choose a code folder, Code in, Claude Code only); live answer How do I let my AI build in a folder on my computer from GilbApp's chat?
Checked against GilbApp d529620.