What connected agents can do
What your AI sees and can change once it's connected — in plain terms, with the guardrails explained.
Once connected over MCP, your AI sees your workspace roughly the way you do: folders containing notes, recordings attached to notes, tasks, screenshots, and transcripts. Here’s what that means in practice.
What it can read
- Notes and folders — list them, read them, search them. Search covers meaning, not just keywords: “the call where they pushed back on pricing” works.
- Transcripts — including exact wording. Your AI can quote what was said, with timestamps.
- Images — screenshots and photos are indexed with descriptions, so “find the architecture diagram from the Acme call” works.
- Tasks and recordings — ask for open tasks, recent calls, or the latest activity in a folder without manually assembling the list.
- Your instructions — workspace and folder instructions are handed to the agent as it navigates, so your conventions follow it around.
What it can change
- Create notes — file meeting prep, research, or drafts into the right folder.
- Edit notes — append action items, update an account plan, fix formatting.
- Move and rename — reorganize folders and notes.
- Delete — remove notes or folders.
The guardrails
- Everything is scoped to your account. A connected agent sees your workspace, nothing else, and only while your token stays valid.
- Edits are grounded in reads. An agent must read a note before editing it, and its edit fails if the note changed in the meantime — no blind overwrites of content it hasn’t seen.
- Deletes have a grace window. Deleted items are recoverable for a while before permanent removal — a mistaken agent delete isn’t instantly destructive.
- You control the key. Access rides on the token from Settings → Connections — treat it like a password, and if one ever leaks, email [email protected] to have it revoked.
How agents refer to things
Agents work with readable paths, the same shapes you see in Banter:
- A note:
/Work/Acme/Discovery call (a1b2).md— the(a1b2)is a short ID so renames don’t break references. - A recording attached to a note:
/Work/Acme/Discovery call (a1b2).md > Call (3f2a).md
You mostly never type these — but when your AI cites one, you’ll know what you’re looking at.
Good first asks
- “List my Banter folders.”
- “What are my open todos in Banter?”
- “Summarize the latest meeting in /Work/Acme.”
- “Create a note in Banter from these bullets.”
Then graduate to the recipes.