New: • Notifications say what happened and answer back. A finished turn shows how long it took, how many tools it ran, what it changed and the first of what it said; a failure shows the error and offers Try again; an approval shows the exact command and Approve asks for Face ID; a question shows its options as buttons. Reply from the Lock Screen to start the next turn, and pull a notification down for a full card. Servers on claude-bridge 1.15 send all of this when the app is closed. • Home is one board. A sentence at the top says what it all adds up to — who needs you, what is working, or that all is quiet — over one vertical list: Needs you, Working, Pinned, Saved, Recent. Every chat is the same row everywhere, with a line only when something happened, and a waiting chat shows the command or question it is waiting on. On iPad and iPhone Duo, Home beside the chat list becomes Today: each waiting question in full, and your quotas. • FLUX 3 Klein paints and edits. The Fast engine now runs FLUX 3 Image Klein: finer detail, lettering that spells what you asked for, and edits that can work from several pictures at once, named image1, image2 in your words. Your ComfyUI machine needs version 0.41 and the new model files; the machine sheet lists anything it is missing. • Watch the app your agent is changing. A Mac running previewd builds it, runs it on a simulator and streams the screen. Open it from the conversation, see each change land, and when a build fails keep the last good picture with the compiler's first error and a button that hands it to the agent. • A video render outlives the app. It is written down before it starts, so quitting or reinstalling no longer loses it: the next launch finds it on the machine and follows it to the end. Changes: • The video button is a light. A ring round its edge shows the machine waking, the queue, how far the render has got, saving, and a clip that landed or a render that broke until you have looked at it. • Stop names its own render, so it never interrupts somebody else's job on a shared machine. • Continue it shows which clip it is continuing, and a clip is saved under its words and seed. Fixes: • The attach button no longer disappears when the model cannot read pictures: Files still attach, and Photo Library says why it is off.

New: the Midgar universe. Open the card at the top of Settings and drift through every app we make. Tap one to see what it does and get it without leaving this app. Apps you already have show a tick, and one tap opens them.

New: • The chat screen is redrawn so every fact has one place. The title stays put while the agent works, with one status line beneath it: Writing, Running Bash, Needs you, Offline, and the clock. • Model and effort, lane, attach and commands are one family of capsules in a single row that fades at the edge, instead of a mix of different buttons. • Finished agents no longer pile up at the bottom of the chat. Only an agent that is working right now asks for room there; a finished one stays at the call that started it. • A turn that failed is a card in the conversation with the server's own words and Try again, and VoiceOver can send it again. • The jump button says what is waiting: Needs you when a question or approval is out of sight, New below when anything arrived while you were up the page. • An empty chat offers four starters; tapping one fills the composer without sending. • With a keyboard, [ and ] jump between prompts and Y copies the last answer. Fixes: • Large text sizes no longer break the permission buttons, the model pill, the composer or the empty chat. • The goal chip no longer has a fixed width, a rule in an answer is a full-width hairline, and a queued message wears a plain glyph instead of an emoji hourglass.

Fixes: • Conversations no longer stutter when a session changes state. The chat list stopped rebuilding itself behind the conversation on every update, haptic feedback no longer waits on the main thread, and opening the app and a chat is quicker. • The app writes far less to your phone: a long answer used to rewrite the whole conversation every five seconds as it streamed, which could add up to more than a gigabyte a day. Changes: • New chats can walk a machine's folders: a folder browser with a path you can tap back along, a filter that narrows as you type, places to jump to, a star on any folder, a count of the chats already working in each, and “Start here” to open the chat in the folder you are looking at. • The Usage screen fits on one screen. Your tightest window leads with a one-word verdict, each provider is one compact card with every window's reset beside its name, providers you can still add share one card, and “How this is counted” is one footer. It reflows at the largest text sizes, and its cards now show in light mode. • The Usage widgets wear each provider's own colour, show a window that has run out as a red mark with “Used up” beside it, show a prepaid balance as one big figure, no longer cut window names short, and on the large size leave out the providers that don't fit instead of running off the tile. • The image studio puts the picture first: the stage fills the screen, shows the machine's own sketch while the picture is painted with a line along its foot for how far along it is, and Save, Share, Copy, Again, Edit this and Animate this sit directly under the finished picture. • The machine's pictures are a strip above the brief, newest first, with the render in flight leading it. The decisions are labelled chips whose values are never cut off, and the machine you are painting on is a pill in the bar. • A third image engine, Qwen Image 2.1 Turbo, paints in eight steps on a machine that has its model; the avoid list and the detail chip stand down for it. • The video forge draws the machine's sketch of the clip as it renders, with a line for each of its two passes, and keeps your clips on a shelf of posters. • Video gains Enhance, using the same helper as pictures, a Start from picture (Photos, Files, the camera, the clipboard or the machine's gallery), Continue it on a finished clip, and Animate this on a finished picture. • Chat is more compact: tool calls, thinking and agent groups rest as one flat line, even a lone tool, a compaction is one line you can tap for its summary, and the space between rows follows what sits on either side of it. Settings › Appearance › Chat density brings the airier layout back. • A question reads as the heading over its answer: what follows it sits at the paragraph gap instead of the gap between turns. • The links in an answer are one line instead of a card each — site icons and names that open in place when tapped, and a single address is simply the link — and the pictures an agent made share a row of thumbnails that open the gallery. • Code blocks no longer leave empty space under their last line.

Fixes: • Opening the app after hours away no longer walks the screen through everything your servers did meanwhile: it reads where things stand once instead of replaying the log, and a burst of list changes lands as a single move. • An answer that lands a table no longer quits the app: the card standing in for a table while it is still being written hands over to the finished table in the same place, and nothing moves under your thumb. • A queued message is sent exactly once, however many places wake up on the same conversation — a phone and its background watch can no longer take the same message twice. Changes: • What you decide about a chat now follows you between devices through your own servers and nobody else's: read and unread, pins, the archive and saved chats are kept by claude-bridge 1.14 and omp-bridge 0.8 (update them in Settings › Software) and show up on every device within a second. A chat open in front of you stays marked read as it grows. • The chat menu gains Pin, Mark as unread and Archive. • Made for iPhone Duo: opened out to its inner display, the workspace shows your sessions beside the conversation, and the conversation keeps to one page while tool output, tables and repository state open on the other. • Held like a laptop, the transcript stops at the fold while the composer, chips and effort rail stay on the base. • Grids use an even number of columns so nothing sits on the fold, and nothing the app draws beside the side bar hides under it.

Fixes: • An answer that lands a table no longer quits the app: the card standing in for a table while it is still being written hands over to the finished table in the same place, and nothing moves under your thumb. • A queued message is sent exactly once, however many places wake up on the same conversation — a phone and its background watch can no longer take the same message twice.

New: • The full model list is the same picker as the control in the message box. Your pinned model-and-level pairs lead it, and every row shows the level a pick would run at, drawn with the same bars and the same carry rule. • The effort strip over the list names the model in use and says whether it changes this chat or the next one; tap or slide to a level, a tick for each. • A star on a row pins that model at the level shown; pins that live on another machine stay listed with the machine that runs them. • Servers are the tabs and providers are the chips under them, and "a pick here starts a new chat on …" is said once above the list instead of on every row. • Opened from Home, from a chat or from a server's default model, the list says what a pick will do before you take it. • Model and effort are one control in the message box, in every chat and on Home: a dot in the model's colour, the model, the effort word and five bars of heat. • Tap the effort half to lift the levels out as a rail and tap one, or slide along the effort half to step through the levels, a tick for each. Swipe the model half to cycle your pinned model-and-level pairs. • Pin a model and a level together. Pinned pairs lead the quick menu, and a pair is offered only where that machine can really run it. • Every provider and local model gets the same treatment: levels such as think and nothink, fast, none or numeric budgets are read, ordered and drawn on the meter. Switching to a model that lacks your level moves it to the nearest cooler one and says so. • Models on other machines are grouped in the quick menu. Picking one aims the composer at that machine and says so, and the machine rows under the target chip show what each will run. • Local and gateway models keep their own name and colour: a Qwen build named after Claude is no longer drawn as Opus. • In the picture lane, holding Generate or Edit (or Home's Send) offers a brief written out by the painting machine's helper. Changes: • Nothing in the list is cut off any more: model names, ids, "used up, resets in …" notes and the machine a model runs on wrap onto another line, and at the accessibility text sizes the controls over the list scroll away with it instead of covering it. • Holding a row in the model list shows a card: its context size, what it reads, the levels it takes, what would become of your level and what a switch re-reads. • From xhigh up, a coloured line runs along the top of the message box, and ultracode wraps it in a rainbow. Fixes: • A row whose provider or limit changed kept its old height and cut its text off; it now resizes. • The last rows of a chat are no longer hidden behind the chips over the message box, and a chat opens on its end and stays there while its rows settle. • The enhance card's Use button reads properly on its glass fill, and the studio's prompt box has a clear button.

New: • "Send these steps to my computer" on the setup screen shares the Tailscale download, the install command and the address command by AirDrop, Messages or Notes, so you can paste them straight into a terminal. • The demo now has a "Connect your machine" card on Home that opens the real setup. • Delegate holds a passing patch for you to read before anything touches your repository. Open the run to see each changed file with its line counts and its diff, then Apply to the tree (unstaged) or Discard. If a chat is working in the same repository, Tailscode names it before you apply. Needs delegate 0.4 on the machine. • Type /delegate in any chat, or choose "Delegate a task…" from the chat's menu, to start a packet on that chat's machine with the task and the repository already filled in. Changes: • Tailscode now asks about notifications after your first agent turn finishes, with a short explanation of the Lock Screen and Live Activity alerts, instead of the moment you first open a chat. • Delegate's board leads with New packet, shows the ladder as one row of tiers with each tier's pass rate, and groups runs into Needs you, Running and Earlier, each with its repository and age. • A run's screen leads with what matters: why it stopped in the verifier's own words, the tier waiting for your approval, or the patch ready for review, followed by one main action. • The new-packet form asks for the goal, the repository (picked from where your chats and runs work) and the paths first; class, verifier and tiers fold into one “How it runs” line you open only to change them. Fixes: • A message waiting in a chat's queue is no longer sent the moment you reopen the chat while its previous turn is still running.

Fixes: • A live chat's status colour on Home no longer flickers every second — it only repaints when the state actually changes. • The clock behind Home's live cards now follows this window's own foreground state, so it keeps counting reliably on iPad, including with more than one Tailscode window open.

The Live Activity now shows what your agent is really doing, and goes away when it should. Live Activity: • Between tools the card says Thinking… again, instead of naming the last tool for as long as the model thinks (with claude-bridge 1.13.1). • A finished turn no longer sits in the Dynamic Island for hours. Starting a card for another chat used to make the server lose track of the first one. • A finished turn's card waits until you open that chat. Having Tailscode open on another screen no longer counts as reading it. • A late update from a turn that already ended can no longer bring its card back as if it were still running. • A turn you stop takes its card down at once. • Deleting a chat removes its card, and turning Live Activities off in Settings removes the ones already showing. • When the agent stops to ask you something, the alert says it has a question instead of asking for approval.

On iPad, Tailscode is now a workspace. And a turn that ends while the app is closed still tells you. iPad workspace: • A sidebar holds Home, Usage, Delegate, Video and Images, your chat lists (All Chats, Live Now, Saved, Archived, one per server) and your pinned chats. • The chat list sits beside it and the conversation fills the rest. One tap hides the columns when you just want to read. • Narrow the window in Split View or Stage Manager and it becomes the iPhone layout, without closing the chat you were reading. • Drag a chat out of the list to open it in a window of its own. A window iPadOS restores reopens the chat it was showing. When the app is closed: • A turn that ends while Tailscode is closed still sends a notification: finished, a question, an approval or a failure. The wait runs straight over your tailnet to claude-bridge 1.13, omp-bridge 0.7 or opencode 2, with no relay in between. • Settings ▸ Notifications shows whether Background App Refresh is on, and each server says whether it can be waited on. Polish: • Tailscode now speaks Polish: onboarding and server setup, Settings, errors, Pro, notifications, the Live Activity, the widget and VoiceOver. What comes from the server — transcripts, terminal output, paths — stays as it is. Fixes: • In the demo, every changed file in the repository view now opens its diff.

Mac servers ask for Full Disk Access once, plus Undo from here, provider waits that explain themselves, faster chats, and Live Activities that stay after a turn. Mac servers: • A server running on a Mac shows whether its agents have Full Disk Access. Without it, macOS stops a turn at a permission dialog on the Mac's screen, where nobody is looking. • One tap opens System Settings at Full Disk Access on that Mac, with claude-bridge shown in Finder beside it. The steps appear on your phone, and the screen notices the moment the switch is on. Adding a Mac server asks once; Not now skips it. Needs claude-bridge 1.12. Fixes: • The model menu lists the server's own models before you have picked one, instead of only Server default. • The update screen no longer crashes when a server restarts, and a restart pressed while the server is already restarting follows it back. • Streaming answers take less work per word in chats with many workflow runs. opencode 2 chats: • Undo from here: touch and hold one of your messages to wind the chat back to it. The messages after it are set aside, the files the agent changed since are put back, and your words return to the message box to edit. Restore brings it all back until you send your next message. • When a provider is rate limited or out of capacity, the chat shows its reason, which attempt failed and a countdown to the next one, with a button to the fix it names. • Quiet lines show when the model or agent changed, when the server restarted and picked a turn back up, and when a background command finished. • A turn cut off by a server restart can be picked back up or let go, and cost and tokens show for every chat. • A message you send appears right away, a chat with an attached file opens again, and stopping a turn or declining a permission no longer shows an error. • Shell commands run from opencode's own terminal appear as they run, files opened by their full path load, and the one-line setup script no longer restarts a server in the middle of an answer. Everywhere: • Hundreds of labels that were still in English now appear in your language. • VoiceOver can reach the buttons on every card in a chat. Faster: • Long chats stay smooth while an answer streams, and answers are written out at up to 120 Hz on ProMotion iPhones. • A chat starts loading as soon as your finger touches it. • Reopening an unchanged chat no longer downloads it again (needs claude-bridge 1.11), and a chat's requests go out together over one connection. Live Activities: • A finished turn stays on your Lock Screen and in the Dynamic Island until you open that chat, saying how it ended, how long it ran and when. • Your next message in the same chat reuses the card, and a card you never open leaves after an hour in the Dynamic Island and four on the Lock Screen. • Running tools show their own icon, updates sent while your phone is asleep appear in your language, and reopening the app keeps the card of a turn still running. Server updates: • When a server has a new version, an Update button appears on the home screen, with what's new and one press to update. • While it updates you see every step with a timer, and afterwards which version the server is on. If an update fails you see which step and why, and can try again or read the log. • Update claude-bridge to get a finished turn on the card while your phone is asleep. Video studio: • A new sound field, and an estimate of how long each render will take on your machine. • The Fine model is gone: it made blurry clips, so the studio renders with the one model built for it.

Pictures from a prompt, written out for you. The image studio can hand a thin brief to a language model on the machine that paints and get back the paragraph Qwen Image was trained on, and you now choose which model writes it: the Enhance menu lists every machine that answered — Ollama, llama-swap, llama.cpp, LM Studio, vLLM — with each model under the name its server gives it and whether it is already in memory. The default is the best writer rather than the fastest, a model you pick is kept, and switching the helper off keeps it filed. The helper is told what the brief alone does not say — which engine paints, the shape you chose (kept, and answered back), how many reference pictures the words address and that their image tokens must stay, and the avoid list — so a 21:9 you picked stays 21:9 and an edit describes the change rather than a new scene. Taking the paragraph puts it in the box where it can still be edited, with your own sentence one tap away. The picture is watched as it is painted. ComfyUI's own sketch after every sampler step lands on the stage in place of the pulse, captioned with the step, and the finished picture cross-dissolves in over the last sketch instead of snapping into place. Your ComfyUI needs --preview-method auto (or taesd) to send sketches; without it the stage shows exactly what it showed before. opencode 2 works. opencode 2.0 rebuilt its whole API — every route moved, every record changed shape, every event was renamed, and a web page answers on every old address — so a 2.0 server used to read as "not an agent server". Tailscode now asks each opencode server which generation it runs and speaks that one: 1.18 and 2.x both connect, list every project's chats in one request, stream answers, tool calls and thoughts as they happen, answer permissions and the new form-style questions, compact, fork and rename, and pick the model, effort and agent per prompt. A server you upgrade under a running app is followed onto its new API without reconnecting. Where a password is typed, the rule is said: claude-bridge, omp-bridge and opencode 2 always want one — BRIDGE_PASSWORD, OMP_PASSWORD, OPENCODE_SERVER_PASSWORD or the one opencode printed when it started — and opencode 1 only if you set one. The add-server form shows it under the field instead of hiding it behind a disclosure, the one-line setup script makes a password up when you gave it none and tells you what it is, and a server that answers and refuses says the same in its own card.

The image studio paints with Qwen-Image-2.1, and it now asks the model for everything it can actually do. The quality engine on your ComfyUI machine paints and edits from one model instead of two: text inside a picture comes out cleaner, an edit follows your reference photo's own shape and keeps the scene it started from. Shape and size are two decisions now. Six frames from square to ultra-wide, three sizes up to the model's native 2K, and both stand down during an edit because an edit takes its frame from what it starts from. One render can work from several pictures, addressed in your words as image 1 and image 2. Cutout paints on transparency. Detail trades minutes for steadiness. Avoid keeps things out, and says that it costs time. The seed rolls by default and holds with one press, which is how you change a single word and see only that word change. This model paints what it is told and invents the rest, so the studio says what a full brief looks like: a short one is flagged under the box, with a frame to fill in, six rules and four worked briefs one press away. And if a small model is running on the machine that paints — Ollama, llama.cpp, LM Studio and anything else OpenAI-shaped — Enhance hands it your three words and gets back the paragraph this model was trained behind. Press again to get your own sentence back.

Two doors to the App Store, in Settings ▸ About: Rate Tailscode opens the review form, and Share Tailscode hands the listing to the share sheet with one line saying what the app is — for the colleague who keeps asking what you are prompting from. The image studio: a tap on the shelf opens the picture and decides nothing; putting one on the stage, rolling it again or starting an edit from it are press-and-hold verbs. The stage starts blank, a start-over button in the bar takes it back there, and the dock sits at the foot of the screen with only its contents riding the keyboard. Tables keep their last row. A table measured before it was on screen came out narrower than its words, wrapped where no fold was planned, and drew its last row past the bottom of its card; cells are now measured once they are in a window. An answer no longer appears twice when it finishes. On the Claude bridge the server could hand a finished reply back under a second name, and the app showed both copies until its next refresh — which was also the flash that sometimes ran through the whole chat at the end of a turn. The app now recognises a reply it already holds, and claude-bridge 1.8.1 fixes it at the source. Effort levels are the model's own on Oh My Pi. A chat on an omp server was offered Claude Code's ladder, ultracode included, whenever the app could not find its model in the catalog — and it could not find any model reached through a gateway, because the server named it door first while the catalog keeps the door and the id apart. The app now finds the model either way, a model the catalog has never heard of gets omp's own levels from minimal to max, and a model that takes no level shows none instead of the last model's "low". A chat started in the omp terminal is named after its first prompt and says which model it runs the moment it appears in the list, without anyone opening it first; omp-bridge 0.5.5 fixes that at the source. A chat whose agent left a command running that will never finish no longer sits in LIVE NOW as BACKGROUND for the rest of the day. The bridge now reads every background shell against the machine and ends one that has shown no CPU time and no output for ten minutes past the budget the model gave it, saying so in the chat. A row carrying work says how long the machine has been at it, work the server found stuck reads STALLED, and a swipe on the row ends the work yourself, with the server's reason when it cannot.

The whole app is now localized — every screen, not just this page. Now in German, Spanish, French, Italian, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, Traditional Chinese.

The whole app is now localized — every screen, not just this page. Now in German, Spanish, French, Italian, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, Traditional Chinese.

The whole app is now localized — every screen, not just this page. Now in German, Spanish, French, Italian, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, Traditional Chinese.

Every chat now shows how full the model's context window is. The band above the prompt box wears a ring filled as far as the window is, with the share first and the token count after it, quiet while there is room, amber as it fills and red when a compaction is close. The number is the footprint of the last request the model was actually sent — what it was handed plus what it wrote — never the turn's whole bill, which for a turn of many tool calls is many times the window; a compaction resets it to what survived. The window itself comes from the server's own catalog, so a local model on your own machine is measured against the limit you gave it, and a Claude model reached through Claude Code against its 200k or million-token window. Tap or click the ring for the whole picture: a large dial with the share inside it, the window banded by what fills it — cache read, cache written, fresh input and the answer — how much is in use, how much is free, which model, where the number came from, what to do about it when the room is getting tight, and Compact one press away.

The whole app is now localized — every screen, not just this page. Now in German, Spanish, French, Italian, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, Traditional Chinese.

The whole app is now localized — every screen, not just this page. Now in German, Spanish, French, Italian, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, Traditional Chinese.

The whole app is now localized — every screen, not just this page. Now in German, Spanish, French, Italian, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, Traditional Chinese.

Ask for a video and watch it being made: an agent renders it on one of your machines while the app follows the work, and the finished film is kept right there. The renderer has its own studio now — one walk sets it up, the model list is a picker you aim into, settings open as a menu, and the chip wears the machine that answered. The composer grows a third lane beside chat and ask, and the send wears the lane it is sending down. When a provider refuses a request mid-turn, the card says what happened and offers the same words again without the pictures.

Usage now lays every provider's rate windows out as readable bars with their own reset clocks, and a restored split layout on the desktop no longer collapses on launch. New App Store screenshots that tell the story properly.

The whole app is now localized — every screen, not just this page. Now in German, Spanish, French, Italian, Japanese, Korean, Brazilian Portuguese, Simplified Chinese, Traditional Chinese.