Can teammates collaborate in Claude Code in real time?

Not natively. Claude Code is a single-user tool: one terminal, one session, one account. As of September 2026 there is no built-in way for two people to work inside the same Claude Code session at the same time, and real-time multi-user sessions are among the most-requested features on its issue tracker (#60082, #27702, #40981, #48828). What you can do falls into four groups: share the setup, share the transcript, run more agents, or share the session itself.

1. Share the setup, not the session

The supported way to "collaborate" in Claude Code is to make every teammate's separate session start from the same place:

• Commit a CLAUDE.md so every session inherits the project's context and rules.
• Check in custom slash commands and MCP server configuration.
• Keep a shared prompt library for the tasks the team repeats.

What you get: consistency. Everyone's agent knows the same things. What you don't: anything live. Two people are still in two terminals, and neither can see, approve or stop what the other's agent is doing.

2. Share the transcript, after the fact

Several tools export or link a finished session so a teammate can read it:

• Claude share links from claude.ai produce a read-only view of a conversation.
• Lore creates a private or public link with the full transcript, tool calls and diffs, readable in a browser or by another agent.
• Session Share (a Claude Code skill) exports the session — history, tool calls, file changes — to a redacted JSON file you can send over Slack or email.

What you get: a good hand-off record and a way to review what happened. What you don't: a way to steer. The decisions were made before the link existed; nobody but the original driver approved anything.

3. Run more agents, not more people

Claude Code's Agent Teams (experimental; enabled with an environment variable) lets one session act as a team lead that spawns teammate sessions, which split work, share a task list and message each other. It is a real multi-agent system.

It is also still single-player: one human owns the lead session and every approval. Agent Teams solves parallelism for one person; it does not put a second person in the loop. If your problem is "my cofounder can't see what the agent is doing," more agents make that worse, not better.

4. Share the session itself

The remaining option is to run the agent somewhere both of you can reach it, with the session as the shared object. That is what Poly is: a browser room with a git workspace behind it, running Claude Code (and Codex) so a team steers the same agent live — multiplayer agentic coding, in the literal sense.

Concretely:

• One timeline — prompts, replies, tool calls, diffs and receipts stream to every member as they happen.
• Approvals from any seat — in manual mode the agent asks before every edit; the approval card appears on everyone's screen and the first person to decide, decides. The timeline records who.
• Agent questions the room answers together — picks are visible live before anyone submits.
• Attribution and per-person billing — every prompt has an author; every turn's credits are charged to whoever ran it. Nobody needs an API key.
• Revert from any seat — every turn is a commit; any member can undo one from its receipt.
• Sandboxed execution — turns run in isolated containers with an allowlist-only network, so a shared agent is not running on anyone's laptop with anyone's credentials.
• A side channel the agent can't read — threads, group chat and polls for arguing about the plan.
• A view-only watch link — anyone, no account, can watch a room build live.

Honest limits: Poly is a hosted product with its own credits (free to start; watching and approving are always free, running the agent is not). It is desktop-browser only for now. And it is a room, not your terminal — if your workflow depends on local tooling the agent must reach, that has to live in the room's workspace.

Which one should you use?

• You want everyone's agent to behave the same → share the setup (1). Do this regardless.
• You need a record of what an agent did for review or hand-off → share the transcript (2).
• You are one person with a big parallelizable task → Agent Teams (3).
• Two or more people need to steer, approve and see the same agent as it works → share the session (4). Nothing in (1)–(3) does this.

Codex too

Everything above applies to OpenAI's Codex CLI as well: it is single-user, its collaboration story is shared configuration, and the same third-party share tools support it. Poly runs Codex in the same rooms as Claude Code, with the same approvals and questions, so a team can mix — each member on the model they prefer — inside one timeline.