Kiteline is a tool for managing coding agents: Claude Code, Codex and others do the work on a computer, and you follow progress, reply and approve permissions from your phone. The desktop side (macOS, Windows, Linux) is open source under AGPL-3.0. No API key, and nothing in the agents' configuration has to change.
In practice the thing that stalls a task is rarely the model failing to write code. It is the agent stopping to wait for a person: run this command, approve this permission, pick one of these options. If you have already left the computer, those few dozen seconds turn into an hour. The last few releases (macOS 3.49.38 to 3.49.40) are mostly about exactly these moments.
Terminal commands run inside the session, not as stray windows
Some commands need a real terminal: code signing has to reach the keychain, long jobs need someone watching the output. Previously each of those popped up another terminal window on the desktop, with no hint about which session started it.
Now the command opens as a tab in the terminal panel of the session that asked for it:
- every tab is labelled with the machine it is on (for example
Mac Studio · zsh), and the line below shows the device, the working directory, or the command that is running and its state; - live output, input and interrupt are available from the phone as well as from the computer; when the command finishes, the terminal exits and the tab closes by itself;
- if the session lives on one machine and the command has to run on another, the terminal still appears under the session that started it;
- on iPhone and iPad two fingers scroll the scrollback, a tap is a click, a long press selects text.
The panel is laid out like VS Code: terminals on the left, a list on the right, a draggable divider, and the list supports reordering, renaming, colours and splits.
Restarting a session keeps the conversation
Settings changed or a new skill added, and running sessions won't pick it up — closing them used to mean losing the context. A session can now be restarted from its … menu, or several at once, including all of them. The conversation is kept and the agent re-reads its settings and skills. Sessions that are working or waiting for approval are not interrupted: they restart after the current turn finishes. Sessions that live in VS Code stay in VS Code; backgrounded ones stay backgrounded.
"Trust this folder?" is handled when you are not there
The newer Claude Code extension asks whether to trust a project folder before it opens a session, and nothing proceeds until someone clicks. Kiteline now clicks it for you in two cases: you have used Claude Code in that folder before, or the project is one you just opened from your phone. Folders you have never used are left for you to decide, with a reminder — the auto-approval covers places you already work in, not every directory.
Two more places where things silently stall
- Backgrounding failed. If Claude Code cannot start in the background on a machine (first-run setup not finished, service unreachable, no permission to read the project folder, folder missing), the top of that session's page now states the reason and what to do, with a retry button.
- A new task typed on the phone filled the input but never sent. On a busy remote server a new session can take over a minute to load; the old 45-second wait gave up and left the text sitting in the input box. Remote projects now wait up to 2 minutes, local ones up to 1 minute, and submit once the session is ready.
Two smaller changes worth mentioning: iPhone and iPad no longer auto-lock while Kiteline is on screen, and pages such as history, projects, usage and to-dos show a skeleton that matches the layout instead of an empty pane while data is being read.
If the agent needs a web page
Also new: when an agent works on a web page — opening it to check the result, testing a flow, filling a form, signing into a dashboard — it uses the browser Kiteline provides, and a small live-preview window appears in the corner of the session showing the page and the command it just ran. Tap it to enlarge. When a step genuinely needs a person (login, captcha, slider), the agent asks you to take over and pushes a notification; you click and type directly in that view, then hand it back. The whole run is recorded and can be replayed at speed, with the commands shown alongside.
Where the data comes from
- sessions, to-dos and usage come from the records agents write on the machine themselves;
- the terminal is a local pseudo-terminal (ConPTY on Windows) — Kiteline only forwards and displays, the agent starts the command;
- anything that needs approval is still pushed to you; it does not decide for you;
- it does not read Claude's credentials and does not call Anthropic's API.
Desktop builds: search hon6/kiteline on GitHub. iPhone and iPad: search Kiteline on the App Store; Android users have a WeChat mini program.
Claude Code is Anthropic's product; Kiteline is an independent third-party tool.
Top comments (0)