What is CMDOP?
cmdop is a messenger staffed by AI employees. You hire a worker in chat by assigning it a role — a persona such as a mail assistant or a technical writer — to a machine you have joined. That worker runs on the computer your company already owns, so it can work with the files, mail, and software you actually use, instead of asking you to move the work into someone else’s cloud.
What you work in is a machine workspace: a conversation attached to a real joined computer, with an optional project folder and an execution engine. It is built for work that continues — not for a one-off administration panel, and not for a marketplace with chat bolted on.
The everyday loop is simple: choose a machine, hire or resume a worker, choose a Project Root when repository context matters, select CMDOP, Claude Code, or Codex as the engine that runs the work, and review the result. Come back later and pick up where you were — the assignment and the conversation both persist.
Each machine still dials out to its relay rather than accepting inbound connections, so a joined computer stays reachable from behind NAT or a firewall with no port forwarding and no VPN. That property is what makes “runs on your own machines” practical rather than aspirational — see How it works.
The idea underneath
CMDOP is an execution-state continuity layer: the thing that keeps live work on a machine coherent while the people and programs attached to it come and go. It is what lets a hired worker’s conversation survive a closed laptop, a different device, or a different person picking it up.
Three properties follow, and they are what actually distinguish the product:
- Execution state is single-homed. One machine owns the state for its work, so there is exactly one place the truth lives. Operators are spread across browsers, terminals, and clients; the state is not.
- A session survives disconnect, and you observe it directly. Close the laptop and the work continues. Come back and you reattach to live state — you are not watching a reconstruction assembled from a log.
- AI is an operator, not a controller. The model attaches to the same session you do and works alongside you inside the same boundary, instead of taking the wheel while you watch.
The blog series treats this in depth.
What runs where
- The
cmdopbinary runs on the machine that has the files, tools, and execution authority. It dials outward to its relay; nothing dials in. - Its embedded relay serves the web console and connects the machine to the rest of its fleet.
- The browser is a client. It renders the model and requests explicit actions — it never becomes the authority for project paths, session identity, or machine permissions.
- CMDOP’s built-in engine uses the CMDOP Router; Claude Code and Codex use their locally installed CLIs and existing sign-in.
CMDOP also exposes operational and interoperability surfaces — relay management, Skills, external MCP tools, and fleet MCP. They support the workspace; they are not the primary work loop.
What it is not
- Not a hosted AI employee that runs in someone else’s cloud. A worker hired through Cmdop acts on your own machine, with the files, mail, and software already installed there — it cannot act on a computer it does not run on.
- Not a hosted computer that executes your project in the cloud. The work runs on your machine.
- Not SSH with a chat window. A session belongs to the machine, so it outlives the connection you reached it on.
- Not an MCP server or Skill marketplace first. Those are distinct surfaces with their own lifecycles. A role is also not a Skill: a Skill is invoked and returns; a role is a persona a contact wears for the whole conversation. See Concepts.
Common questions
What is Cmdop in one sentence?
Cmdop is a messenger staffed by AI employees: hire a worker in chat, assign it to a machine you have joined, and it works there — on the files, mail, and software you already have — inside a durable conversation that survives disconnects.
Does Cmdop require SSH or port forwarding?
No. SSH can still be useful for direct shell work, but Cmdop’s normal path does not require an SSH daemon, port forwarding, or a VPN. The machine agent dials out to its relay, so nothing has to dial in.
Where does the work actually run?
The work runs on the joined machine that owns the files, tools, and session. The browser, terminal, SDK, and MCP clients are operators attached to that machine’s live session; they are not separate places where execution state is copied.
Does hiring a worker send my data to the cloud?
The files, mail, and records a worker touches stay on the machine — nothing is bulk-uploaded. When you use CMDOP’s built-in engine, the text of the task itself is sent to the CMDOP Router for inference, the same way any AI chat product calls a model provider; it is not a claim of fully offline or zero-cloud operation. You can also point the router at a model you run yourself.