AI agents now run for hours or days — but the session still belongs to one laptop. Costeer puts any coding agent in a live room your whole team can join: watch it work, steer it, hand it off — with real permissions, and a permanent record of who did what.
Tell us about your team and what your agents are doing — takes 30 seconds.
What Costeer does
Everyone in the room sees the same live stream: the agent's plan, its tool calls, the diff taking shape. Joining is a link, like a Figma file.
One person drives. Teammates queue suggestions the driver sends with a click. Dangerous actions route to whoever's account is allowed to approve them — a viewer can't green-light a shell command.
Pass the wheel across accounts and close your lid; the session keeps its full context — it moves to the next person's machine and carries the conversation with it. Every room persists as a searchable record: who steered, who approved what, on which branch.
Where the agent actually runs
The room is shared. The agent is not — it runs locally, in each engineer's own repo, on their own Claude account. Costeer relays the session; it never holds the keys to it.
Not a policy — a shape. The connection is outbound only, and the protocol has no field that could carry a credential, so there is nothing on our side to store, leak, or subpoena. Your code stays local too, unless you push it.
A prompt is dispatched only to the current driver's machine. Nobody's work runs on a teammate's account, and the record shows whose credential executed each turn — the audit trail your finance and security teams will both ask for.
Anyone can propose a command; it runs only if the owner of that machine approves it. Room permissions don't override that — holding the highest scope in the room still doesn't let you run anything on someone else's laptop.
Today the agent starts from one command in a terminal. A desktop app is in progress — same model, one install, with approvals as native prompts instead of a browser tab.
Why now
“Think Google Docs / VS Code Live Share, but for a Claude Code chat.”
Early access
If your team runs long agent sessions — background refactors, overnight migrations, on-call handoffs — we want you in the beta. Request access and tell us what your agents are doing.