Every device opens on Sessions, the home page; an explicit URL still wins. The bottom bar is Sessions · Approvals · Tasks · Settings · More (Terminals joins it while a terminal is open), and Approvals carries the count of what is waiting for you. Session status uses four words everywhere — Working, Needs you, Idle, Ended — and "Needs you" means an approval or permission prompt, never an agent quietly at its prompt. Blank shells with no project sit under their own Scratch terminals heading, apart from Sessions and projects, and both lists keep the search, scope and grouping controls. Make a project on a scratch card promotes it without moving its files or restarting its terminal. Agent cards put Chat first.
Start an agent opens with a Folder, an Agent and one toggle, Ask before risky actions (on for a new install), and one Start button; the name, group, launch profile, model, worktree isolation, sandbox, start mode and first message wait behind More options. Choose… browses the machine's folders; starting in a folder that is not a project yet adds it to your projects. Only agents found on that machine are offered, the rest under More agents…; with none installed, Try a demo agent runs a scripted stand-in. The chosen folder and machine are spelled out beneath the picker, and the launch summary next to Start says whether the agent will ask. This browser remembers the project you picked last (a deliberate new empty folder included) and lists recently opened projects first, in Start an agent and in Terminal's new-terminal menu. A creation that fails is not remembered as recent.
Two quick shells never read the same: each scratch card is titled with its own scratch folder, shows that full path so it can be read or copied, and keeps ✎ Rename for a name you recognize. Renaming changes the card only — the folder, its files and the terminal stay where they are — and the name survives a reload. Cards nobody renamed are titled from their folder; stored rows are never rewritten behind your back. Terminal remains one tap away. Chat preserves an unsent draft on the device, shows connection state, and offers working-file changes. Needs you collects pending approvals, setup/task failures and work ready for review. A failed refresh is shown explicitly. The chat box sends on Enter (Shift+Enter adds a line); on a touch-only phone Enter adds a line and the Send button sends — Settings → Basics → Enter in chat changes either.
In Terminal, session names have their own full-width row. The switcher, new terminal, search, menu and navigation controls use a separate row. Swipe sideways on the terminal body to move between open terminals, including full-screen apps that enable terminal mouse reporting. Vertical drags still scroll; long press selects text (see The phone terminal) and pinch changes text size. The terminal fits the visible viewport when the phone keyboard opens, including browsers that overlay the layout.
↺ Restore lists closed, archived, exited and restart-interrupted sessions, grouped by project with a search box. Each row shows why it closed, its last message and what one tap will do — usually Resume, which continues the saved conversation. Other agent… continues it with a different agent or model. Ending a session from its card shows an Undo toast for ten seconds, and sessions a restart interrupted get a banner that restores them in one tap. Sessions a restart brought back on their own are listed in a dismissible Relaunched N sessions notice. A card whose agent quit to a shell prompt says Stopped and offers ↻ Revive; the terminal view shows the same prompt. See Restoring sessions.
Switch agent or model offers discovered models, saved provider profiles and favorites. Switching saves a fresh handoff and starts the destination with that context. It keeps the original session running. The progress display distinguishes saving from starting, reports failures, and lets you retry. If the new session already exists but its terminal cannot open, Open new session retries only the attachment. Go back opens the predecessor; it does not restart it. This is a context handoff between agents, not a claim that their native conversation formats are interchangeable.
Favorites are saved per browser. A favorite for a deleted provider is marked unavailable; editing a saved provider uses its current configuration. Keep API keys in a protected server-side credential source or launcher, never in model names, command arguments or browser storage. A launcher profile is specific to the targets where its command and credentials are installed.
- Select while it runs. Hold a finger on the output: the word under it is selected in the live terminal, and the same finger drags the selection further (the view scrolls at the top and bottom edges). The agent keeps running and output keeps arriving. The bar at the top offers Copy, Whole lines, Open file / Open link when the selection is one, and Done; a tap anywhere clears it. Tools → Pause view still turns the whole buffer into native selectable text.
- Tap to open. A tap on a web address opens it in the browser (the phone's
own browser in the Android app); a tap on a workspace path, including
path:line:colas compilers and test runners print it, opens it in the file viewer. OSC 8 hyperlinks work the same way (web addresses, andfile://links into the workspace). Paths outside the workspace are not offered. - Your key row. ✎ at the end of the row (or Tools → Customize key row) reorders and removes keys, adds the ones not shown (⏎, ⌫, Del, ⌕ find, …), any Ctrl/Alt/Shift combination (Ctrl-R, Alt-B, Ctrl-←, Shift-Tab …), and your quick commands as one-tap keys. A quick-command key points at the command itself (stored on your Lectern, the same list as ⚡ and Settings), so editing the command changes the key on every device, and a project's commands only appear in that project's terminals. Holding an arrow, ⌫, Del or Page Up/Down repeats it; a drag along the row scrolls it without pressing anything. The arrangement is stored on the device; Reset restores the default.
- Pull to refresh at the top of Sessions or Tasks fetches everything again (not the whole page).
- Swipe a session card right to approve what it is waiting on, left to archive it (a live session asks first, since archiving stops it). Every archive gets the usual Undo toast. A short swipe springs back; a mouse never swipes, and the same actions stay on the card's buttons.
- Menus are bottom sheets on a phone: a card's More opens from the bottom edge within thumb reach, over a dimmed page that closes it.
- Offline. The lists every home screen starts from (sessions, tasks, projects, machines, pending approvals, live views) are kept on the device after each successful load. Opening the app with no network, or with the relay unreachable, shows them at once under Offline · showing what Lectern said N min ago; the marker goes as soon as a request gets through, and Retry asks again. The copy lives in the page's own storage for that Lectern, so it works the same in a browser, the installed web app and the Android app (which has no service worker); a 401 or forgetting a pairing clears it. Actions still need the connection.
- Haptics. A committed swipe, a pull that will refresh, a long-press selection and the start of a key repeat give a short tick: natively in the Android app, through the Vibration API in browsers that allow it.
Under each pairing QR code (Settings → Phone & devices), Copy link and Share…
send the same one-time link by message or email instead of scanning it. The
https link works in any browser and in the installed web app. Open in
Android app is its lectern://pair?… twin, and the pairing page itself, opened
in a phone browser on Android, offers the app too. The app only fills in its
pairing screen from a link; nothing is paired until you tap Connect. See
android.md.
The 🎙 buttons can transcribe on the Lectern host with whisper.cpp instead of the browser's own speech recognition. It is the only option in the Android app (its WebView has no speech recognition) and a choice anywhere else for keeping audio off a browser vendor's service. The phone records 16 kHz mono WAV, sends it to your Lectern only (through the relay when paired over it), and the text lands in the box for you to check; nothing is sent on your behalf.
Setup on the host: install whisper.cpp so whisper-cli (or whisper-cpp) is on
PATH, and put a ggml model where its download script puts them (for example
~/.local/share/whisper.cpp/models/ggml-base.en.bin), or set
LECTERN_WHISPER_BIN and LECTERN_WHISPER_MODEL. LECTERN_WHISPER_LANG
(default auto) and LECTERN_WHISPER_THREADS are optional. Nothing is
downloaded for you. Choose the engine in Settings → Notifications → Voice
input; Automatic uses the device when it can. Transcription needs a
signed-in person (like deciding an approval), runs one recording at a time, and
takes up to five minutes of audio (three minutes per recording from the app).
Voice mode's continuous conversation still uses the browser's speech
recognition.
When an approval is decided anywhere (another phone, the desk, the terminal, or it expires) Lectern sends every subscribed device a withdrawal for that approval's notification. The Android app removes it. A browser closes it too; because browsers require every push to leave something visible unless the app is on screen, when it was the last notification showing it is replaced, silently and under the same tag, by one line saying what happened ("Approved on another device"). ntfy-topic and Discord messages cannot be withdrawn.
Native keyboard and gesture coverage, limitations, and the nightly audit are in Mobile terminal testing. The Android fixture is disposable and never types into a person's running agent.
Chat for a live session defaults to structured cards, not a dump of the
tmux pane. The server locates the session's own Claude JSONL or Codex rollout
file — the same identity-verified reader the saved-conversation picker
already uses, reused rather than reimplemented — and decodes it into typed
turns (assistant/user text as light markdown, a collapsed Thinking entry
for private reasoning, and one card per tool call) instead of the flattened
display string the plain-text history reader produces. The phone polls
GET /sessions/{id}/conversation/live and only asks for what was appended
since its last look; Terminal text is one tap away in the same controls
row, and is what a session falls back to automatically if its log can't be
found (a shell-tracked session, an agent without a reader, or a sandboxed
session) — nothing more to do, the toggle itself simply isn't offered.
Every tool call collapses to a title and, on a phone, one tap away: Read/
Edit/Write/MultiEdit show the file path and, for an edit, a real diff (the
same renderer the session review diff uses); Bash and Codex's exec_command
show the command and its output; Grep/Glob show the pattern; WebFetch/
WebSearch show the host or query; TodoWrite renders as a checklist; a
sub-agent Task shows its description; an MCP tool's mcp__server__tool
wire name becomes "server · tool"; Codex's apply_patch gets the same diff
treatment, split per file. A tool this registry doesn't specifically know
still gets a real card — name plus collapsed, pretty-printed arguments —
never a raw JSON blob.
Approvals are graduated, and one card shows them everywhere — the session
card, the Approvals page, the chat and Needs you: Allow once (key Y)
decides just this call; Allow for this session (A) tells the broker to
stop asking for this same tool, or this Bash command's first word, for the rest
of the session, entirely in memory and forgotten when the session ends (a task
attempt's approval has no session, so there it is Always allow in this
project); Deny… (N) opens a note that is sent back to the agent as
the hook's denial message. The
approval's command or diff renders through the same tool-view registry as a
chat card, never JSON.stringify(approval.input). Claude and Codex
interactive sessions share one PermissionRequest hook path end to end, so
"Allow for this session" behaves identically for both; it has nothing to do
with Codex's separate app-server driver, which only ever backs a scheduled
task attempt, never an interactive chat.
A Now strip sits at the top of Sessions: one chip per live session (working / waiting / done / error) and, when any exist, how many approvals are waiting — a tap on either jumps straight there. It is the PWA analogue of a native app's Live Activity, since a browser tab cannot keep one on the lock screen. The app icon's own badge (where the platform supports it) mirrors the same count — set from the page on every refresh, and bumped by the service worker itself on a push, so it stays roughly right even between page loads; it clears the moment nothing needs you.
Wherever there is a text box for a reply — the chat composer and the Needs-you "deny with reason" field — a mic button appears if the browser supports the Web Speech API and stays hidden if it does not. Interim words show up as they are heard and are replaced, never duplicated, once recognized; nothing is ever sent on your behalf; you still tap Send.
Below the composer in a live session's Chat view, 🎧 Voice mode turns the same free Web Speech API into a hands-free loop instead of just dictation — no vendor keys, no metered API. Tap it once; nothing listens until you do.
- Talking to the agent. While on, it listens continuously and shows the live transcript. Pausing for about 1.5 seconds, or ending with "send it" or "over", sends the utterance — after a 2-second "Sending… tap to cancel" window, so a misheard phrase never reaches the agent unreviewed.
- Hearing it back. New agent output is read aloud, sentence by sentence, as it arrives. Terminal noise (ANSI colors, box-drawing borders) is stripped first, and a fenced code block is announced as "Code omitted." rather than read character by character. Starting to talk immediately interrupts whatever is being spoken (barge-in) — you are never stuck waiting for a reply to finish before you can jump in.
- Spoken approvals. A pending approval is announced — "Claude wants to run <command>. Say approve or deny." — and answered by saying approve / allow / yes or deny / no / stop, alone or with "please" (a sentence that merely mentions the word, like "don't approve that yet", is never mistaken for the decision). The decision is always echoed back, spoken and on screen ("Approved." / "Denied."), before it takes effect.
- Voice commands: "interrupt" or "stop that" sends the same Escape key as the composer's Interrupt button; "read that again" repeats the last thing spoken; "exit voice mode" turns it off.
- Settings (voice, speaking rate, language, auto-read on/off) are stored per device and reached from the ⚙ link in the panel.
- Fallback and capability handling. Where continuous background listening
is not available — notably iOS Safari, including inside an installed PWA
on some versions — voice mode falls back to push-to-talk: hold the mic
button, speak, release. A browser with neither
SpeechRecognitionnorspeechSynthesisnever shows the toggle at all; it explains why in its place instead. A screen Wake Lock is requested while voice mode is on (and tolerated if the browser refuses it), so the screen does not sleep mid-conversation.
Notifications for the same session replace each other in the tray instead of stacking (grouped by session, or by approval, or by kind for a broadcast like a failed check). A session waiting on you gets Open terminal and Reply actions alongside the usual tap-to-open — Reply opens the chat with the composer already focused. iOS Safari not yet added to the Home Screen gets explicit numbered steps in Settings → Notifications, not just the one reason sentence every other unavailable case gets.
The terminal's Tools menu on a phone carries one more, deliberately low-profile item: Report keyboard layout, which captures the exact viewport/keyboard numbers this device is reporting — for pasting into a bug report when the phone keyboard covers the prompt in a way no emulator has reproduced (see the keyboard-coverage history in Mobile terminal testing).
A session card's ⋯ menu has Add to favorites / Remove from
favorites, and a starred card shows ★ before its name. Starring is in the
menu rather than a button of its own so a card keeps at most three actions. The
sessions list has a Favorites filter, so a phone can show only the starred
sessions. The star is the same one the terminal client sets
with * (see terminal-client.md), and it
survives the session ending. Opening several favorites at once is a terminal
client feature; on the phone you open them one at a time.
A claude session can also be opened from claude.ai/code or the Code tab of the Claude mobile app: tick Also open in the Claude app (Remote Control) when starting it, or use ⋯ → Enable Remote Control on a running one. A card with it shows a small ⌁ by its name, and ⋯ → Open in Claude opens it there. It is the same session, so approvals and messages from Lectern keep working. See remote-control.md for requirements, the one-time consent and what happens on restore, fork, switch and account swap.