Fleet MCP tools

Cmdop’s MCP server is a fleet interface for a local assistant. It operates joined machines through Cmdop; it is not a general local filesystem or shell interface. The surface has two layers: the fleet core — listing, asking, running one command, reading files and directories, and Machine Care — plus a set of projected capabilities that give the registered assistant the same higher-level actions a Cmdop conversation has. Each tool targets a machine you name. The exact set is versioned with the Cmdop binary, so treat cmdop mcp tools as the source of truth for a given install.

The fleet core

Tool	What it does
`list_machines`	Lists your fleet with each machine’s online status and OS.
`ask_machine`	Sends a natural-language task to a machine’s own agent and returns its reply.
`run_command`	Runs one shell command on a machine and returns its output and exit code. A non-zero exit is a result, not an error.
`read_file`	Reads a file from a machine, paging a large one rather than dumping it.
`list_dir`	Lists one directory on a machine, non-recursively.
`care_status`	Reads durable health facts for a machine, with their freshness.
`care_diagnose`	Runs one bounded process or startup diagnostic on a machine.
`care_storage_inventory`	Reads the validated Cmdop-owned storage inventory for a machine.

Projected capabilities

Beyond the core, the fleet surface also projects Cmdop’s higher-level actions so a registered assistant can do more than list, ask, and read:

Capability	What it does
Browser control	Drives the target user’s own signed-in Chrome to act on the web (the `browser_*` tools).
`web_search`	Runs a web search and returns results.
`generate_image`	Generates an image from a prompt.
Screen capture	Lists screen targets and captures the screen (`list_screen_targets`, `capture_screen`).
`publish_site`	Publishes a static site to a public Cmdop address.
`machine_info`	Reports a machine’s host facts.

These tools reach other machines across the relay. They are not a local shell — the assistant already has its own shell and file tools for the machine it runs on.

Every tool’s target accepts a hostname, display name, unique prefix, or machine UUID. Ask the assistant to list machines first, then name the exact target: a known machine can still be offline or require its own connection PIN, which is passed as an argument to ask_machine and run_command.

Agent-to-agent boundary

When a local assistant calls ask_machine, it is asking the target machine’s own agent to do work inside that machine’s boundary. That is agent-to-agent command execution at the product level: one assistant reaches another joined machine through Cmdop’s fleet tools, not through SSH, a VPN, or a copied local shell.

Keep the target explicit. List machines first, choose one online machine, and keep its reply separate from replies from other machines. Fleet tools do not turn several machines into one shared shell or one transaction.

Inspect your installed tool set

Tool names, descriptions, and availability are versioned with the Cmdop binary. Check them on the joined machine that registered the assistant:


cmdop mcp tools

The command reports the fleet core — listing machines, asking a machine’s agent, running one command, reading a file or listing a directory, and reading Machine Care facts or running bounded care diagnostics — alongside the projected capabilities above, where those tools are available.

Treat the command output as the source of truth for the tool list in a given installation. MCP clients can present those tools under client-specific names.

Access boundary

Each request uses the joined machine’s existing Cmdop fleet access. Target machines can still require a connection step or be unavailable. The MCP server does not create a separate fleet account, expose a local network endpoint, or give an assistant direct access to the caller’s local shell or filesystem.

Common questions

Which MCP tools let an assistant run commands on another machine?

Use list_machines first, then ask_machine or run_command against the exact target. ask_machine asks that machine’s own agent to do work, while run_command runs one bounded shell command and returns its output and exit code.

Can one Claude Code or Codex session operate several servers?

Yes. The assistant can list joined machines and address one online target at a time through Cmdop’s fleet tools. Keep each server’s reply, PIN, timeout, and error separate; the tools do not turn multiple servers into one shared shell.

Are these tools the same as SSH?

No. SSH opens a shell on a reachable host. Cmdop’s MCP tools address joined machines through the relay, so the target can be behind NAT and does not need an inbound SSH port for the AI-agent workflow. SSH can still be the right emergency or direct-administration surface.

To register the assistant in the first place, see Claude Code and Codex. To give a Cmdop agent tools from an outside server instead, see External MCP tools.