One agent is a tool. Several is a team.
Vect runs the agent you already use in a window built for it, and seats several of them in one working directory when one is not enough: each given its own job, each reading what the others wrote before adding its own, and the room saying so when two of them touch the same file. On your own machine. No account, no server, nothing uploaded.
You would not put one person on all of it.
Nobody staffs a project with one person and no reviewer. A room is the same arrangement, owned by you: the work split, each part read by something that did not write it, and every step of it in a record you can go back through afterwards.
Work too big for one sitting is a queue when one agent does it and a morning when three do the parts at once. Each of them only has to hold its own part.
The one that wrote a claim is the worst one to check it. A second reads what the first produced against your own sources and says where it does not hold.
Ask two of them for the same thing in different ways, then pick, rather than taking the first answer because it is the only one you have.
A drawing set, a grant application, a pile of interviews to read: work that has parts, a reviewer and a date is work a room is for.
Together, without overwriting each other.
Run two agents on the same work and they normally trample each other: both edit the same file, one saves over the other, and you find out later. A room is one working directory they are both inside. Each is given a different job, each is told to read what the others have written before adding its own, and when two of them touch the same file within moments the room says so rather than leaving it for you to find in a diff next week.
A session starts with one agent, which is the right number most days. Add a second to a conversation already in progress and the room forms around it, keeping everything said so far.
How rooms work →chapter three by friday. nothing goes in that sources.bib can't back.
· drafts started on chapter-03.md · teammates read it first
outline agreed. starting on the cold open, flagging claims as i write.
the outline dates the storm 1907. the archive says 1909. flagging it before it spreads.
Fix the date everywhere it appears? Three files.
· you allowed · edits changed 3 files · outline.md released
chapter three is in: 2,400 words, every claim pinned to a source. over to you.
· you denied · nothing changed · the flag stays on the thread
One agent drafts a chapter while another checks every claim in it against your sources and a third keeps the outline honest, each reading the others before it writes.
Send several readers at a pile of papers at once. They divide it up, and what one finds is on the table before the next one goes looking for it again.
Two agents on one feature, splitting implementation from tests, reviewing each other, and flagging it when they both land in the same file at once.
One agent is a fine way to use it
Most days it is you and one agent, and that is the whole app: your terminal, your files and one conversation in a window that is nicer to sit in than a terminal tab. Bring the agent you already pay for. A second one is a button, not a migration.
Read the docs →One window for the whole project
A tab holds the conversation, a terminal and a browser pane, with your files a click away in the side panel. Drag any pane onto the edge of another and it splits there. It opens as just the conversation, and the rest appears when you ask for it.
Read the docs →It remembers what you were doing
Every conversation Claude Code and Codex have had on this machine is browsable and searchable here, including the ones you started in a terminal before installing Vect. Work you want to come back to goes on a board, where a card can carry its own branch.
Read the docs →Agents can drive the app
vect mcp serve hands any agent Vect itself as tools: running things in the terminal you are looking at, opening a page in the browser pane, starting other agents, and reading and writing what the project remembers. Spawning, running, opening and remembering all land in the audit trail as they happen.
Say it instead of typing it
Tap the mic and talk, and it types what you said. It can read replies back to you once you turn that on. Both speech models run on this machine, downloaded the first time you use them, so dictating a paragraph does not send your voice anywhere.
Read the docs →It works where you already are
vect chat opens the same agent in whatever terminal you had open: the same memory of the project, the same permission prompts, the same record afterwards. A room of several agents can be started from a script the same way, and what they did is waiting in the window when you look.
Your phone as a second screen
Scan a QR code and your phone is paired to the machine running your agents. Read along as they work, reply, and answer their questions from another room. It runs in the phone browser, over your own network, so nothing leaves the machine.
Read the docs →A room keeps its own record. Who joined, which files each of them touched, and where two landed in the same one, on a timeline you can read afterwards instead of reconstructing from a diff.