Alelyon app 0.7.0, the patch notes.

Alelyon app 0.7.0 17 October 2026

• Closing Lattice no longer risks a crash when the Models view finished checking this machine's models just before the close.
• My page shows the top of your page as others will see it (banner, photo, name, title, tickers, biography and totals), drawn again as you type. Change photo and Change banner sit on it, and Remove banner shows only when there is one.
• Name, Title and About you say how much of each is used, and a title the service would refuse is said in place before anything is sent. Discard changes puts back what was saved; Enter in a field, or Ctrl+Enter in About you, saves.
• Background: Add a position and Add a school open their forms on demand, with "I hold it now" and a bad month or year said before it is sent; Delete asks first. Skills are chips, each with how many endorsed it and an x.
• Close my page moved from Profile to Privacy.
• Lattice: a request to Claude Code or Codex that is sent just as the agent ends now says the agent stopped at once, instead of waiting out its timeout and saying the agent did not answer in time.
• Your own AI agent can drive Alelyon: Settings, Agents, "Let my agent drive Alelyon". A program you run on this PC, such as alelyon-cli's MCP server for Claude Code, Cursor or Codex, reads the app's chats and runs, and once you allow it in the banner that asks, opens sections, starts and sends chats, opens web pages and starts runs; everything it does shows in the window. It listens on this computer only, with a new secret each time Alelyon starts, in a file only your Windows user can read. Off until you switch it on.
• alelyon-cli gains the app's tools (app_status, app_chats, app_chat_read, app_runs, app_go, app_chat_new, app_chat_open, app_chat_send, app_open_url, app_run_start) and `alelyon-cli app`; its MCP server runs without a developer token for them.
• Numbered editors beside the chat: 1 is yours, and each agent that works on files gets the next, with its own tabs and a dot while it works; following opens an agent's files in its own editor, so the one on show stays put; Alt+1 to Alt+9 picks one.
• Lattice is chat first: the chat is the page, in a reading column, and the editor a panel beside it (Ctrl+E or </>) that opens on the files the agents read and edit while it follows them, with a Working on list of what each agent touches; hidden during a turn, it stays hidden until the next.
• The chat box at the centre, your work at the side: before the first message the empty chat is centred on the chat box, with the greeting above it and the suggestions below, which go away once the first message is sent; the side bar opens on Work, where the file list was: what the agents work on now (each agent, the step it is on and its files, each opening in the editor), then the to-do list, the changes waiting, the stack's suggestions, the checklists and the tools, each a list that folds to its head, so that nothing stacks over the chat box.
• A new look for Lattice's chat: your messages as bubbles on the right; answers as text under the agent's mark, a size larger, with Copy and Regenerate beneath; each turn's work one line ("Worked for 1 min 12 s · 9 steps · 2 files edited") that opens to its steps, open while the agent works and folded once it is done, with what waits for you always in view; the to-do list, checklists, stack suggestions and changes waiting as chips above a round chat box, whose bottom holds images, Agent or Ask, the project, the folder, the file in view, the model and Send; earlier topics that fold to one line, and side-list sections (Groups, Topics) that fold away, a group's arrow showing its topics; and, with no chat yet, a greeting with four ways to start.
• The tools, a press from the chat box: a Tools chip opens Agent or Ask as cards and a card for each thing the agent can use (files, commands, its browser and your desktop with their switches, MCP servers, plugins, skills and commands), each saying what it asks first; the empty chat greets you by name and the time of day; and the side list shows your projects, each opening to its topics.
• Work goes to the group already doing it: No group in the side list is a fresh page, and the message sent there goes to the group already doing that work, the page following it to that chat (also when that chat is busy and the message waits its turn); work a group's task did goes to the chat that started the task, a group's name and plan count toward what it does, and work no group does starts a new group.
• Plan limits on the harness, ready for paid plans: tasks at once, helper agents, schedules and the project manager follow the signed-in plan (Free: 2 tasks, 4 helpers, auto mode; Pro and Team: 6 and 12; Max: no limit). Switched off until purchases open; nothing changes for anyone today.
• Connector discovery: find_connectors searches a built-in catalog of local MCP servers; suggest_connector adds one to the user's MCP settings only after their yes.
• Desktop depth in auto mode: right, middle and triple clicks, drag, held keys, zoom, several displays, the clipboard (reading it asks every time) and opening a program from the Start menu.
• web_fetch reads a public web page as text without the browser (redirects re-checked, this PC and private networks refused, redacted); view_image shows the model an image of the folder, and read_pdf a page of a PDF rendered by Windows, when the model can see.
• Steering helpers: spawn_agent runs a helper in the background (wait_agent collects its answer), send_agent steers a working helper or asks one that answered again, close_agent ends one, and spawn_agents fans one helper per task over a list (four at a time); a helper still working when its turn ends is stopped.
• The agent's browser opens the user's own app on localhost (dev servers it started; other ports after a yes), works in tabs, and uploads a folder's file into a page (always asked; secret files never sent).
• Notebooks by cell: read_notebook shows a notebook's cells and outputs; notebook_edit stages a replaced, inserted or deleted cell for review.
• Schedules and watches: the agent schedules prompts (once or repeating, as a new topic or a wake-up of its chat) after the user's yes, run only while Alelyon is open (missed runs reported, not run late), listed with Cancel on the Tools page; watch_command waits for a background command's output.
• The agent and your terminals: read_terminal reads a terminal's last lines and run_in_terminal runs a command in a new terminal of the agent's own; each asks you every time, and nothing is typed into a terminal you opened.
• Walkthroughs: write_walkthrough saves the turn's pictures of the agent's browser or the screen, in order with captions and notes, as an HTML artifact; long artifacts show the start of their source.
• Team seats in Settings, under Account: a Team plan's owner sees the seats taken and the people on them, invites a person by username and removes one; a person sees whose team they are on and can leave it, and accepts or declines invitations waiting for them.
• Lattice's Models, Foundry, Training Studio and System tabs read at a glance: status tiles and compact tables up top, two columns on a wide window, explanations behind a ? on hover, and rarely needed parts closed until opened; the planned-feature cards left the Models tab (they are on the Roadmap in Settings).
• Research, the Models panel and your own page in Pages need far less scrolling: the OpenAlex key and the agents' setup fold into one line each, a subject's facts sit in two columns under one header line, the Research hub lays its lists side by side, the Models panel shows what is in use above tabs for local files, hosted models and endpoints, and your page splits into Profile, Background, Privacy and Applications tabs. Explanations moved behind a ? beside what they explain.
• Pages read at a glance: titles take one line, explanations sit behind a '?', wide windows use two or three columns, Sinai's Auto configure and Machine parts open on a press with their state on one line, Data shows a row beside its grid, the Words library is a row a transcript, and every planned feature is on one Roadmap in Settings instead of a card on each page.
• Pages and tabs in windows of their own. Drag a page's place off the rail (or a tab of Lattice, Transcription or Compute down off its bar) and let go over the desktop or the page, press the pop-out button beside the window's buttons, middle-click, or use the palette. The page or tab moves to its own window, marked where it was; a press there brings it forward. Back, or closing its window, returns it, and its keys work in its window.
• Builders on Pages: a Products tab ranks the day's and the week's launches by support and lists the newest products, your own and those you follow; a product's page shows its logo, stage, links, platforms, technologies, screenshots, makers and stars, with its updates, launches, feedback board and reviews; its keepers edit it, list makers, add a logo and screenshots, and invite the people who ask to test it; the composer posts an update or a launch About a product you keep; posts show their product, launches carry Support, feedback carries votes and a status its makers set, reviews their stars; organizations get a Products tab, search finds products, seven new notifications, Launching today beside the feed, and code blocks between ``` lines.
• alelyon-cli, a command line and MCP server for developer tokens: your own AI agent (Claude Code, Cursor, Codex, Lattice) reads and acts on Alelyon Pages with only the access the token allows; the token is read from standard input and kept sealed for your Windows user.