MobSessiondocs

The orchestrator, the mob, and the agent

The core mental model behind MobSession: three roles, who controls what, and why the agent runs on your real machine.

MobSession turns one Claude Code session into something a whole team can watch and steer at once. There are three roles to keep straight: the orchestrator, the mob, and the agent. Every session has exactly one orchestrator, one agent, and a mob of any size.

The orchestrator

The orchestrator is whoever runs mobsession start. The session lives on their machine, in their working directory, using their installed tools and their Claude subscription. Nothing runs in a remote sandbox.

The orchestrator holds final control. They can:

  • Drive the session from their own terminal, typing prompts like any normal Claude Code session.
  • Interrupt the agent at any moment with ESC, stopping work mid-stream.
  • Set the working directory and decide which tools are available.

The mob steers, but the orchestrator can always take the wheel.

The mob

The mob is everyone who joins the session link. No install: open the link in a browser and you're in. The mob can do two things:

  • Watch live. The agent's output streams to everyone at once.
  • Submit prompts. Anyone can send a prompt that steers the agent, just like the orchestrator does from the terminal.

That's what makes it a mob rather than a screen share: one source of truth, and anyone can add direction.

The agent

The agent is the driver, a real Claude Code session doing the actual typing, reasoning, and tool calls. Not a transcript or a replay, the live thing doing the work.

  • One shared output. Whatever the agent does streams to everyone at once. No per-person view drifts out of sync.
  • One shared input. Prompts from anyone, terminal or browser, flow into the same agent. Terminal and web are two doors into the same room.

Who controls what

  • Orchestrator: owns the machine, directory, tools, and subscription. Drives directly, interrupts with ESC. Final say.
  • Mob: watches live and submits prompts to steer. No install.
  • Agent: executes. Takes input from terminal and browser, streams one output to everyone.

Control flows downhill from the orchestrator, but ideas flow in from everyone.

Why the real machine matters

The agent runs on the orchestrator's actual machine, not a sandbox. That has real consequences:

  • The work is real. The agent edits real files, runs the real build, uses the real tools. When the session ends, the changes are already on disk, not stuck in a throwaway environment.
  • The context is real. The agent sees the project exactly as the orchestrator does: same dependencies, config, and state. No gap between the demo and the codebase.
  • Control stays local. Everything runs under the orchestrator's account and subscription, so they can always interrupt, redirect, or stop from their own terminal.

Mobbing on a sandbox is watching a toy. Mobbing on the real machine is doing real work together, with one person in control of the environment.

Next steps