Does Claude Code Remote Control use more tokens than the terminal?

No. Remote Control does not add a model call. A message sent from your phone lands in the same local session, on the same model, with the same context, and is billed exactly like the same message typed in the terminal. What changes is your behaviour: from a phone you tend to send more, shorter turns, and every turn re-sends the whole context. That is where a remote session can end up costing more, and it is the same effect you get by typing more in the terminal.

What does Remote Control store on Anthropic's servers?

While a Remote Control session is connected, the transcript (your messages, Claude's answers and the tool activity) is stored on Anthropic's servers so that your phone, your browser and your terminal show the same conversation and the session can reconnect. Code execution and file access stay on your machine. The transcript is storage, not inference: it does not spend tokens by itself.

How do I check the token cost of a remote session from my phone?

Type /usage in the Claude app or on claude.ai/code: it is one of the commands that work from a connected device. It shows the current 5-hour and 7-day windows as percentages. Note the numbers, send a few messages, run it again. The delta is what those turns cost, and it is the same delta you would get from the terminal.

Does Codex have a Remote Control like Claude Code?

Half of it. codex-cli 0.154.0 ships an experimental codex remote-control command: start and stop an app-server daemon with remote control enabled, and pair to print a short-lived pairing code. The documentation describes it as meant for managed remote-control clients and SSH remote workflows, and does not name a phone app or a web page you connect from. There is no equivalent of the Claude app tapping into your terminal session yet.

Can I drive Codex, Cursor or Copilot CLI from my phone today?

Yes, with a different mechanism. The AgentsRoom mobile app is a remote for the terminal itself: what you type on the phone is sent over an end-to-end encrypted relay into the local terminal of the agent, whichever of the 14 supported CLIs it runs. Nothing goes to a provider that was not going there already, so the token overhead is zero by construction, and it works for Claude Code sessions too, with or without Remote Control on.

Does Claude Code Remote Control Use More Tokens? And Does Codex Have One?

Two questions people type into Google since Anthropic shipped Remote Control: does driving a Claude Code session from your phone cost more tokens, and is there an equivalent for Codex. Short answers: no, a turn is a turn wherever you type it, and yes but only half of it exists. Here is what Remote Control actually is, how to measure the token question yourself in four commands, what codex remote-control does today, and how a mobile remote that never talks to a provider changes the math.

September 16, 2026

Since Anthropic shipped Remote Control in Claude Code, two questions have started showing up in our Search Console, on pages that do not answer them: "does claude remote control use more tokens" and "does codex have a remote control like claude". Both deserve a straight answer. Here is what we checked, on Claude Code 2.1.273 and codex-cli 0.154.0, and how you can check it yourself.

What Remote Control is, in one paragraph

Remote Control connects the Claude app on your phone, or claude.ai/code in a browser, to a Claude Code session that keeps running on your machine. Three ways in: claude --remote-control (or --rc) starts an interactive session you can also drive remotely; /remote-control (or /rc) turns it on inside a session that is already running; claude remote-control starts a server mode that waits for connections and shows a QR code. The session, the filesystem, the MCP servers and the tools stay local. What leaves your machine while a device is connected is the transcript, stored on Anthropic's servers so that every connected surface shows the same conversation and the session can reconnect.

It needs a claude.ai login on a Pro, Max, Team or Enterprise plan. An API key does not work: the CLI refuses with "Remote Control requires claude.ai subscription auth". And the local process must keep running: close the terminal and the session goes offline until you resume it.

Where the tokens go

A Claude Code session spends tokens in one place: each turn sends the context (system prompt, files read so far, previous exchanges, your new message) to the model, and the model answers. That is true whether the message came from your keyboard or from your phone. Remote Control does not add a model call. It adds a transport, and a copy of the transcript on Anthropic's side. Storage is not inference. There is no second session, no summarisation pass on the phone, no "mobile model".

So the honest answer to "does it use more tokens" is: not per message. The same message costs the same wherever you type it.

What can change is how many messages you send, and how long the session gets. From a couch, with a phone keyboard, people send shorter turns and more of them: "ok", "now the tests", "and the other file". Each of those is a full turn, with the whole context re-sent. Ten short turns cost more than one long one carrying the same instructions, because the context is paid ten times instead of once. That is not a Remote Control tax, it is the normal cost of a chatty session, and you would pay it in the terminal too. The mitigation is the usual one: batch what you want into one message, and let auto-compact do its job on long sessions.

One more place people worry about: push notifications and the "Still working, check in from your phone" link Claude Code shows on long turns. Those are notifications about a turn that is already running. They do not start one.

Measure it in four commands

You do not have to take our word for it. /usage is one of the commands that work from a connected device, so the whole measurement can be done from the phone.

In the terminal, start a session with claude --rc and connect the phone.
From the phone, type /usage. Note the 5-hour and 7-day window percentages.
Send three ordinary messages from the phone, the kind you would send from the terminal. Wait for the three answers.
Type /usage again. The difference is what those three turns cost.

Now do the same three messages from the terminal, in a fresh session of similar size, and compare. The two deltas move together. If your remote delta is bigger, look at the transcript: you almost certainly sent more turns, or the session was already deeper. /context from the terminal shows how much of the window is occupied; that number, not the device, is what drives the cost of the next turn.

The same reasoning answers a neighbouring question we also see, "does claude voice mode use more tokens": the voice layer transcribes and speaks, the model sees text either way, and a spoken turn is billed like a typed one.

Does Codex have a Remote Control like Claude?

Half of it, as of codex-cli 0.154.0.

codex remote-control exists, marked experimental. It has three subcommands: start launches the app-server daemon with remote control enabled, stop halts it, and pair prints a short-lived pairing code (with --json you get pairingCode, environmentId and expiresAt). OpenAI's documentation says it is meant for "managed remote-control clients and SSH remote workflows", and explicitly not a replacement for codex app-server --listen when you build your own client.

What the documentation does not name, at the time of writing, is a client. There is no page saying "open the ChatGPT app and pick up your terminal session". The server side is there, the pairing is there, the phone side is not documented. So if the question is "can I continue my Codex terminal session from my phone the way I do with Claude Code", the answer today is no, not with OpenAI's own tooling.

The other way to do it: remote the terminal, not the session

There is a second approach to the same need, and it is the one AgentsRoom takes. Instead of asking each CLI to grow a remote mode, you remote the terminal the CLI is running in.

The AgentsRoom mobile app pairs with the desktop app over a relay that is end-to-end encrypted: the relay forwards bytes it cannot read. What you type on the phone is sent into the local terminal of the agent, and the terminal output comes back. The agent does not know a phone is involved. That is why it works for Codex, and for Cursor, Copilot CLI, Antigravity, Grok Build, Mistral Vibe and the rest of the 14 CLIs AgentsRoom launches, on the day the CLI is supported, without waiting for a remote feature on their side.

It also settles the token question by construction. Nothing is sent to a provider that was not going there already: no transcript copy, no extra session, no relay-side model. A message from the phone is a message in the terminal. The relay costs bytes, not tokens.

The two mechanisms coexist. A Claude Code session started inside AgentsRoom can run with /rc on: Anthropic's Remote Control gives you the Claude app's conversation view with the transcript synced, and the AgentsRoom remote gives you the raw terminal, the other agents, the backlog and the dev servers of the same machine. Most days the terminal view is enough. On a long turn, the "check in from your phone" link is a nice thing to tap.

What to remember

Remote Control does not bill a message differently. A turn is a turn. Watch the number of turns and the depth of the session, not the device.
/usage works from the phone. Measure before and after, that is the whole method.
Codex has the server half of a remote control, experimental, with a pairing code, and no documented phone client yet.
A terminal remote works for every CLI at once and adds zero tokens, because it never talks to a provider.

If you want the terminal remote, the remote agent control page shows what the phone can drive, and Claude Code Remote covers the Claude specific setup. If the real question was how much you have left this week, how many tokens do I have left in Claude explains the two screens to read.