Lab · Design concept

Claude and Codex are single-player. Here’s a team layer for both.

Both desktop apps are excellent at running many agents for one person. Neither lets a team see that work, own it, hand it off or talk about it where it lives. These are 14 concept screens for the missing layer: an inbox, shared projects, tasks with owners, handoffs and messages anchored to the work.

October 2026 · Independent concept · Select any screen for full size

01 · The gap and the model

Many agents for one person, nothing for the team.

I went through what each app ships today. Parallel sessions, worktrees, routines and scheduled tasks are all there. Shared projects, tasks with an owner, messages between people, an inbox of what needs you, handing work to a colleague and delegating an approval are not. Claude’s sessions can already message each other, but only inside one account.

The fix fits both apps because both already have the spine: a project that starts sessions, and a list of what is waiting. The team layer adds people to the project, owners to the work, and messages that live next to it.

A comparison table of Claude Code desktop and Codex in the ChatGPT desktop app across ten capabilities, marking each as ships, partial or missing.
The gapWhat each app ships today for the things a team needs. Both run many agents for one person; neither runs a team.
Five cards describing the Member, Project, Task, Message and Session primitives, a flow showing a message becoming a task, a pull request and a review, and where the layer lands in each app.
The modelFive primitives fit both apps: Member, Project, Task, Message, Session. Agents are members too, always with a human owner.
02 · Inside Claude Code desktop

An inbox, shared projects, tasks, handoffs, a channel and follow.

Six screens in the Code tab. The Inbox joins the sidebar next to sessions and routines; projects gain members, a task board and a channel; a session gains Hand off and Share.

The Claude Code desktop sidebar with a new Inbox, team projects and teammates with presence, and an inbox of approvals, handoffs, mentions, task assignments and review requests.
Claude Code · InboxOne list of what needs you across every project: an approval a teammate routed to you, a handoff, a mention, a review request, findings nobody owns yet.
A shared project with four people and two agents: the project conversation on the left, and an Overview grouped into waiting on you, waiting on others, running and done.
Claude Code · Shared projectA project with members. The Overview shows everyone’s threads by state, and transcripts open only where their owner shared them.
A task board with backlog, assigned, running and needs review columns, cards owned by people and by Claude, and a task detail panel with its activity log.
Claude Code · TasksOne board where people and agents each own a card. Claim it, dispatch it to Claude, or hand it off; the log says who moved it.
A hand off dialog over a Claude Code session, choosing what travels to the teammate, with a preview of what she sees in her sidebar and inbox.
Claude Code · Hand offHand a session to a teammate. The summary, the branch and a note travel; permissions, credentials and side chats never do.
A project channel where a person asks a session for status, the session replies with a diff excerpt, a handoff is logged, an approval is forwarded, and a mention picker lists people, sessions and routines.
Claude Code · ChannelMessages anchored to the work: @people, @sessions and @routines, a diff quoted in the thread, an approval forwarded to whoever owns staging.
A teammate's shared session opened read-only, with the live transcript on the left and the diff with an inline comment on the right.
Claude Code · FollowWatch a teammate’s shared session live, read-only, and leave comments on the diff. Steering needs her say-so.
03 · Inside a Codex-style shell

The same layer, built on Activity, Team Tasks and Space.

Four screens in an original shell that follows the structure OpenAI documents for Codex in the ChatGPT desktop app. Teams already exist there; the change is giving work an owner and connecting shared pages to the desktop.

A dark Codex-style desktop shell with Activity filtered to the team, grouped into needs you, needs someone, running and ready for review.
Codex · ActivityActivity becomes the team inbox: needs you, a “needs someone” queue anyone can claim, what is running, and what is ready for review.
A team task table with owner, runs as, trigger, status and linked columns, and a new task form choosing a Codex owner and approval routing.
Codex · Team TasksTeam Tasks get an owner: a person, a Codex chat, a Dot or the team account, with approvals routed to a named human.
A share dialog over a Codex chat with per-person access levels, the option to push the worktree branch, and an expiry.
Codex · Share a chatShare a chat at view, comment, steer or transfer, with the worktree branch pushed so reviewers can check it out.
A shared plan page where an @Codex mention assigns a task to Marco, and his desktop Activity showing the assignment waiting for him to accept.
Codex · Space bridgeAn @Codex mention on a shared page becomes an owned task with a stop condition, waiting on its owner’s desktop. Status flows back to the page.
04 · Rules and build order

Eight rules that keep a team layer from becoming a surveillance layer.

The screens only work if people trust them. These rules are what every screen above depends on.

  1. Agents are members, not tools.Every session, routine and persistent agent has an owner, a face and a state. Agent-authored messages are always labeled with the owner’s name.
  2. Sessions are private by default.Sharing is opt-in per session and scoped: summary, read-only, follow, steer or handoff. It is time-boxed, revocable, and logged where the owner sees it. Side chats never leave the owner.
  3. One owner per task at a time, human or agent.A handoff changes the owner and carries the context. Nothing is copied silently, and the log says who moved it.
  4. Every message anchors to a work object.A session, task, diff, pull request, artifact, page or project. Free-form chat stays in Slack.
  5. “Needs you” is personal; “needs someone” is the team’s.The inbox is per person; the queue is per project. Nothing waits on nobody, and an unowned item can be claimed by anyone with access.
  6. Approvals can be routed, never bypassed.A forwarded approval shows who asked, who answered, in which session, and it expires. The agent still runs inside its owner’s sandbox.
  7. Permissions never travel between people.A handed-off or accepted task restarts in the receiver’s permission mode with their connectors and credentials.
  8. Aggregate by default.Teammates see titles, owners, states, pull requests and tasks. Transcripts and diffs open only by share. A private session is listed as private, never summarized.
  1. Shared projects and the task listMembers, shared instructions, library and routines; tasks with one owner, linked to sessions and pull requests.
  2. The inbox and routed approvalsOne per-person list of what needs you, with approvals a session owner can forward to the person who owns the resource.
  3. Handoff and messages anchored to sessionsHand a session to a teammate with a summary and branch; @-mention people, sessions and routines where the work is.
  4. Follow, steer, and standing team agentsWatch a shared session live, ask to steer, and give routines and persistent agents a team owner with a sign-off person.
A phone screen showing the inbox with a routed approval, a handoff, a mention and a review request.
PhoneThe inbox follows you. Approve once, continue a handoff on your Mac, or ask a session to answer a teammate.
On the phone

The inbox follows you.

Approvals, handoffs and mentions are the parts of the layer that cannot wait for a laptop. On the phone, the inbox is the whole product.

Independent concept. Not affiliated with or endorsed by Anthropic or OpenAI; product names belong to their owners. The Codex screens use an original shell, not a recreation of OpenAI’s interface. All people, repositories and numbers are synthetic. The comparison reflects each vendor’s public documentation as of October 2026.

Building AI tools for teams?

The hard part is ownership, not the model.

If your product runs agents and your customers work in teams, this is the conversation I’d start with.

Let’s talk → More in the Lab