Phone apps for coding agents sit next to your code and secrets. Where does the data go?

A new link in the chain

Coding agents such as Claude Code and Codex run on your computer and work through tasks on their own, stopping now and then to ask permission. A growing set of companion apps lets you follow that work from your phone: remote viewers, notifiers and approval apps.

As dev.to reported, these tools sit right next to your code, your conversations with the agent and any secrets the agent hands back to you. That makes a simple question worth asking: when your phone talks to your agent, where does the data travel?

The post uses one app, Kiteline, as its worked example. It is an independent third-party tool, not made by Anthropic. Its desktop side runs on Mac, Windows and Linux and is open source, meaning anyone can read the code. The phone app is on the App Store, with no Android version yet.

What the app reads, and what it leaves alone

Agents already save their sessions as files on your machine. According to the post, Kiteline reads those files and nothing else. It does not touch the agent's login, its API keys (the passwords that let software call an AI model) or its configuration. The agent keeps talking to its model exactly as it did before.

There is a trade-off. Because the app only sees what each agent writes down, what you can do from the phone differs from agent to agent. The desktop app lists, for each one, which actions work and which do not.

Direct first, relay only as a fallback

On the same Wi-Fi network, the phone connects straight to the computer. Away from home, the app first tries a peer-to-peer connection, where the two devices link up directly. Only if that fails does it fall back to a relay, a middleman server that passes messages along.

The post says the content is end-to-end encrypted in either case. That means it is scrambled on one device and unscrambled only on the other, so the relay handles unreadable ciphertext and cannot read the conversation.

Devices pair by scanning a QR code on the computer, which swaps the keys the two sides need. A device that has not been paired cannot connect.

Giving secrets their own place

Agents sometimes hand you sensitive items, such as an app ID, a gateway URL and a private key. Dropped into a chat reply, these are hard to find later and easy to leak, for example in a screenshot.

Kiteline gives agents a "key info" command. The phone then shows a separate card with copy buttons, with sensitive fields hidden until you tap them. You can save the card to favorites.

Sharing follows a similar local-first idea. A session can be shared as a read-only link. Your computer serves the page on the local network, the viewer can read the conversation and steps but cannot send anything, and you can revoke the link whenever you like. The post says nothing goes through a server.

Proof the work got done

The post treats trust as having two halves: where your data goes, and whether the agent really finished the job. Kiteline ships a few agent skills aimed at the second half.

• A browser test report with a screenshot for each step.
• Task stages, such as "2/4 testing", shown in every list.
• A separate reminder for things only you can do.

None of this answers the privacy question by itself, but it shows what these apps can add beyond simple notifications.

Reading this with a careful eye

The post describes Kiteline's own design, and these are the app's claims, not the result of an independent audit. Open-source code on the desktop side lets outsiders check it, but the post does not say anything about what the relay stores or for how long.

That is the useful lesson for any companion app. Ask the maker what is relayed, what is stored and for how long. Check whether messages are end-to-end encrypted, and avoid sending a client's private repository through a tool you have not vetted.