Release notes

Files attached to a message now show as a rail of pages you can open, with the rest behind a +N, and Dropstone is told which files came with which message, so "same again for these" answers the newest sheet rather than the one before it. An artifact can be renamed in its panel and saves under the name it shows, and every file a reply produced can be exported together as one zip from under the last one.

Asked "did you use my memory on that reply?", the CLI used to say no even when it had just applied one of your lessons, because the recalled memory only existed for the turn it arrived on. Each turn now keeps a record of which memories it was given, so Dropstone can name what it drew on and which lessons it was handed but did not need. See what Dropstone has learned explains how to ask.

Also in 1.0.49 and 1.0.50: a server error no longer ends a turn (Dropstone retries three times with growing waits while the loading line stays on screen), tool rows lead with what they did (Update(path) with "Added 31 lines, removed 3 lines" underneath), and a recorded lesson leads with Learned (rule) or Learned (fact).

Select text and a pill offers Marker, Note, Quote and Copy. The selection has pointers on each end that snap to whole words and can be dragged; an existing highlight can be recoloured, set to none, or removed from its card. Quote puts the passage in a tray above the prompt that says whose words they are and offers follow-up questions written for that passage. Any reply can be forked from its action row: the conversation up to that reply, in a chat of your own.

Dropstone addresses people by handle in a shared chat, and a question asked in a thread carries the room's handles so it knows who is who. Highlight a passage, quote it, and ask about it and Fork a chat cover the details.

Attaching a Word, Excel or PowerPoint file used to fail with "functionality not supported". Every document format is now converted to text before it reaches the model, and each one shows as a named chip in the prompt: [Word 1], [Excel 1]. A scanned PDF with no text layer is read with vision, up to 30 pages.

The transcript got out of the way: a running command streams behind its row instead of filling the screen (click the row to watch it), and file changes show their first ten lines with the rest behind a click. You can now switch projects mid-session with ctrl+o or /open, and take back a message you queued behind a running turn by pressing up with an empty prompt. A session signed in to an organization reads its policy before the first turn, so a restricted model is explained up front rather than failing partway through.

Replies, comments and highlight notes typeset LaTeX, so a formula is drawn rather than printed with its dollar signs, and a wide equation scrolls on a phone. Highlights and comments now work in your own chats, not only shared ones, and your notes reach Dropstone as yours. A question asked in a thread with @dropstone is answered as a short margin note, streamed into the thread. A quiet plan island at the top of a new chat says which plan you are on.

A chat could be handed to someone as a frozen copy. It can now be handed to them as the chat: they open it, see it as it happens, and, if you allow it, ask Dropstone in it. Share then Invite people makes a link per chat, as an editor or a viewer, and a People button in the header shows who is in it and who is here now. One person asks at a time; everyone sees the reply.

Around that: every account has a handle, so @name addresses one person and tells them they were named; typing @ offers the list; select a passage to highlight it in one of five colours with a note, and everyone in the chat sees the mark; a bell on the left lists where you were mentioned; and naming @dropstone inside a comment thread gets an answer in the margin rather than a new turn. Dropstone's own comments carry a verified tick. The Work together section covers all of it.

A table the conversation kept coming back to, an outreach board, a task list, a comparison filled in over several turns, was re-pasted into every reply. A tracker is that board as a thing. Dropstone creates one when a table is going to be kept up, updates only what changed, and you open the current state from the Tracker button instead of scrolling to the newest copy.

Signing up now asks for a username, and you can change it from Settings. It is what @name mentions in shared chats resolve to.

Settings then General opens with a box for standing instructions: how long, how formal, what to skip. It is saved to your account and sent with every message, on every device. Replies are also shorter by default: one or two sentences for a simple question, and length grows with the question.

Connectors live in one table now, built in and remote alike. GitHub, Notion, Linear and any remote MCP server you add by URL sit beside Google and SmartThings, with a status for each. A shared chat is now readable by anything that follows the link, so a pasted link shows a real preview, and /share/<id>.md returns the transcript as text.

A turn thinking for nineteen minutes looked identical to one that had died. The loader now shows a live clock and the model's own words about what it is working through, the reasoning trace streams while it happens and folds into a timeline afterwards, split per pass, and a long think gets a card that says so. Durations read the way a person says them: 45s, 1m 20s.

Underneath, a turn can no longer end empty because a connection dropped: long generations are followed straight from the server with no time ceiling, a turn whose connection died is continued rather than lost, and Stop actually stops.

Memory got sharper too. Recall now reads the recent turns together, not just the last message, so "yes, do it" still finds the note about the thing being agreed to, and Dropstone checks memory against what it is about to do, not only what you typed. When it thinks a lesson would help your team, it offers to share it and you decide; it never shares on its own. Memory in teams and organizations explains the levels.

An organization gives a company one Dropstone account structure: members, teams, centralized billing, shared memory and enterprise sign-in, managed from the dashboard's organization console. Chat honours the organization's policy on every request, so a model or connector an admin has not allowed is refused with a clear message rather than a generic error, and your seat shows beside your name.

Chat also moved onto the same frame as the dashboard, so moving between your chat and your organization console does not feel like two products. Dropstone for teams and organizations and Shared memory across a team have the details.

Pro moves to a new model that is both faster to reason and far cheaper to run. A Pro turn now costs about what a Fast turn costs, so you get roughly nine times more Pro work out of the same weekly allowance. Pro also understands images natively, so screenshots and mockups go straight to the model that answers. Fast and Heavy are unchanged, and the 1.6 lineup has been retired from the model list.

In the CLI, a change that spans several files now lands in one step, and each mode carries its own icon beneath the prompt. A billing error was corrected in the same release: Fast turns had been charged above cost, and are not any more. Choose a model describes the tiers.

The input sits in its own frame behind a marker, with the agent, model and effort chips on a line below it, and the messages you have sent are drawn in the same frame so the conversation reads as one thing. Tips now appear beneath a running turn, where there is a moment to read one, and far less often.

Asking about something Dropstone has learned now returns the memory that actually answers, not an unrelated one that looked similar. In our own testing the right memory went from coming first in half of cases to all of them. Every so often, when a turn completes, Dropstone asks how it went and takes a single keypress; nothing is sent if you have turned off Improve Dropstone.

Since August 28, new sessions start at Medium reasoning effort instead of Max: on our own measurements the tasks that completed did so at the same rate either way, and Medium is faster and uses less of your allowance. Max is still there with /effort. Since August 31, commits Dropstone makes for you carry it as a co-author, and the license ships alongside the binary (dropstone licenses).

The Faultline page shows what each connected repository's dependency check found. The check runs in your own CI and reports back; the page never triggers anything. Three outcomes rather than two, because "nothing breaks" and "a fix is proposed" are both normal, and a page that only surfaced problems would make every update look like an incident. You choose which repositories the dashboard shows.

Every plan's weekly allowance went up, and chat and the CLI show the new numbers. How usage and limits work explains how allowances are counted and when they reset.

Every reply was held to a fixed reasoning budget, and because the thinking and the answer share it, a hard problem would reason almost to the limit and have little left to reply with. That cap is gone: Dropstone asks for everything the model you picked can give, and a single piece of work can run for hours, or days, with Dropstone following it to the end.

In chat, model calls now run as background jobs on the server, so a turn that reasons for longer than a connection can stay open is still answered. The CLI also adapts its depth to the problem, reaches for tools to verify before concluding something is unknowable, and says when it is uncertain instead of overstating confidence. Errors became readable: a short message with what to do next and a link to the status page, instead of internal detail.

/share in the CLI publishes a session to chat.dropstone.io as a read-only page: a summary of the files it changed at the top, then each turn's reasoning and tool calls, live while the session is still running. /shares lists everything you have published and revokes any of it, including sessions shared from another machine. A shared session shows only the files that session changed, when it ran, and on what.

In the same release the loader says what the model is actually working through, a short recap appears under each finished turn, runaway repetition is caught mid-response, and a message sent while the agent is working waits its turn instead of derailing it. Share a conversation covers both kinds of share.

The SDK reaches 1.0. It uses your account's memory, so something you taught Dropstone in chat or the CLI is available to code you write against it. Dropstone SDK and Using the API with your plan explain what it is for and how access works.

An attached file used to exist for exactly one message: you would get a good first answer and be told on the next turn that no file had come through. Attachments now stay in the conversation for every later turn. Spreadsheets can be read cell by cell and edited in place, with formulas computed, and Word documents can be edited too; a download card appears for any file Dropstone changes. Replies reveal word by word as they stream, and a reply that would have come back empty now says why. Upload and work with files has the formats and limits.

On the dashboard, deleting your account now also cancels your subscription in the same step, and accounts created with email and password need a verified phone number before the dashboard opens.

Chat no longer has a sign-in of its own. You sign in once at dropstone.io, with Google, GitHub, or email and password (new the day before), and chat, the dashboard and the CLI share that account. A prompt typed on the homepage before signing in is carried through sign-in instead of being lost.

The homepage and the dashboard were rebuilt, and Contact became a chat that takes your request and replies in place. Create an account and sign in covers the options.

Before asking the model for a single token, chat used to make five round trips in a row, measured at fourteen seconds of a twenty-eight second turn. They now run from a cache and in parallel, so the first token arrives in a fraction of the time. The reasoning shows as a timeline, grouped per pass, with the tool calls that followed each thought.

When Dropstone writes something you will send yourself, an email, a post, a message, it now appears as a draft card with two genuinely different versions. Nothing is ever sent from the card: the button opens your own mail or social client with the text filled in, and you press send. Select any text in a reply to quote it into your next message. Chat also knows about the CLI now, and will tell you when a task is better done there.

Press ctrl+o in the CLI to change the folder Dropstone works in, without quitting and starting again. Click the mode, model or effort in the footer to change them instead of remembering the shortcut.

Every tier reasons harder at every effort level, and a new max level sits above xhigh, in violet so you can tell when it is on, in both chat and the CLI. Usage credits can now be switched on from chat's Settings as well as the dashboard, and the limit dialog offers a top-up when that is the right answer. How usage and limits work explains credits.

Tell Dropstone how you want something done and it keeps that as a standing rule, applied to every task, instead of only recalling it when the new work looked like the conversation it was learned in. /memory lists everything it has learned, rules separately from facts. Press x on an entry, or say "forget the rule about X", and it forgets, and only then. Ask what it remembers and you get a real answer read from your memory rather than a guess.

The day before, it started telling you when it is drawing on something it learned, and recording a correction the moment you make one rather than at the end of a piece of work. How Dropstone remembers you is the place to start; Rules and facts explains the two kinds.

Word, PowerPoint, Excel, OpenDocument, EPUB and RTF files can be attached and read directly, and PDFs work on every model. Attaching a document no longer freezes the session, and a stalled request gives up and tells you instead of waiting indefinitely.

Fast, Pro and Heavy all moved to new models. Heavy now runs a much more capable, and much more expensive, model, so it draws down your weekly allowance considerably faster; the CLI says so when you select it. Use Heavy for long-horizon work and stay on Pro or Fast for everything else.

The model you pick now sticks between runs and is no longer overridden when you open an older conversation, screenshots work on every tier, and switching between Build, Accept All and Plan takes effect immediately, even partway through a task. Dropstone can also offer to change mode when the current one does not fit.

You can buy a Dropstone subscription for someone else. They receive it by email and redeem it on their own account; an unclaimed gift is reminded about and then returned to you as credit. Gift subscriptions explains how it works.

In chat, a sent image now shows as the image, in the bubble, and opens full size, instead of a "image.png" chip.

Chat, artifacts, the model picker and settings all work on a phone now. The sidebar becomes a drawer, panels become full-screen sheets, and dialogs get a gutter. Dictation moved to a new speech-to-text engine, and replies can carry coloured highlights.

Plans can be paid for in Indian rupees, the dashboard's Overview now leads with a "Start building" section, and a Max subscriber can move up to the larger Max tier with the difference prorated. Pay in your currency and Compare plans have the details.

In the CLI, Dropstone now keeps track of how its own work turned out (how many turns, whether any tool call errored, whether the run completed) so its self-learning is grounded in what happened. Metadata only, and none of it if Improve Dropstone is off.

Connect SmartThings from Settings and Dropstone can see your devices, tell you when one is offline, control them, and set up watchers on your home. Calls Dropstone makes for you now show as one card per call, call audio in a shared conversation is opt-in, and phone numbers are redacted everywhere a conversation is shared. Two days later, watchers gained a one-time mode alongside repeating, and recordings can be deleted from Settings.

The free plan's weekly allowance was doubled. When you do reach a limit, chat now shows a dialog with how much you used, when it resets in your local time, and what to do, and a strip above the composer while you are over. Models that need a paid plan are marked in the selector rather than failing. How usage and limits work explains the reset.

Skills are packaged instructions Dropstone can follow: browse the catalog, install one, and enable it per chat from the composer's Skills menu. Create your own, attach documents to it, publish it privately or publicly, and see installs and usage in the Creator Studio. Verified authors carry a mark, skills have versions and ratings, and the marketplace is searchable. Skills covers creating and installing them.

News is a For You brief written for you each day, date-anchored and country-aware, alongside a Trending feed drawn from several sources; go deeper on a story, track it, or tell Dropstone you want less of it. Stories you have already seen are left out of the next brief.

Watchers let Dropstone keep an eye on something for you and tell you when it changes; set one up from a chat, and manage them in Settings. You can also edit a message you already sent and resend it.

Hand a long job to a background agent and keep working; it reports back when it finishes. Dev servers, builds and file watchers can run in the background without freezing your session or timing out. Press the down arrow to open a panel that lists every running agent and watch its live logs, and x to stop one cleanly. Background agents follow the mode you are in: unattended in Accept All, strictly read-only in Plan.

Discovery proposes candidate solutions, runs each in a sandbox, and applies only the changes a verifier proves are measurably better, for example faster code that still passes its tests. Press Tab in the CLI to switch to it. Chat gained the same loop, with code verified in a sandbox before it is kept.

Dropstone now records lessons from working with you, studies sources you give it (a URL, a PDF, a document) and keeps what it learned with the source attached, and recalls all of it across sessions. Auto-recall brings what you have already studied into a task before it searches the web. A researcher sub-agent studies a topic in depth and saves what it learns, with citations.

Chat and the CLI share the same memory from the first day: one memory, attached to your account. This is the start of what the Memory section describes; Memory across chat, CLI, and SDK explains the sharing.

Attach a spreadsheet, a deck, a PDF or a document and ask about it. Upload and work with files has the current formats.

Fast, Pro and Heavy moved to a new model lineup after an evaluation of candidates. The CLI's home screen became a welcome panel that greets you by name and shows your model, plan, quick tips and what is new in the latest release, and /changelog lists every released version. A badge lights up when your weekly allowance is temporarily boosted.

Your sessions can now help improve the open models Dropstone runs. It is on by default and you can turn it off at any time from the dashboard or chat settings. Dropstone learns from task outcomes and which edits you accept or reject; enterprise and commercial data is always excluded. Privacy controls has the switch and what it covers. The dashboard also gained a Changelog page with a CLI tab, fed live from each release.

Dropstone Chat launches at chat.dropstone.io. Conversations are saved to your account, with the same Fast, Pro and Heavy models and the same effort ladder as the CLI. From the first day: web search, projects with their own files, artifacts for the things Dropstone makes, voice mode and read-aloud with word-by-word highlighting, press-and-hold dictation, phone verification, Google sign-in, mail and calendar, and calls Dropstone can make on your behalf with a recording kept in Settings. Get started with Dropstone Chat is the tour.

The billing page was rebuilt: a transactions list, a clear path to move down a plan, and an Add credits button that is always there. Usage credits can be switched on so that a funded balance carries you past your weekly limit. Settings gained an API keys section for the public API, and the Max plan replaced the Teams plan name. A Trust page describes where requests run and how data is handled. Your dashboard walks through the pages.

dropstone login opens dropstone.io, you sign in there, and the CLI is signed in. The install scripts moved to dropstone.io/install so one command installs on macOS, Linux and Windows. Install the CLI has the command.

General availability. The CLI reaches 1.0 with the three-tier lineup (Fast, Pro, Heavy), interactive and one-shot modes, MCP server support, agent and skill plugins, and self-updating binaries for macOS, Linux and Windows. The following ten days of releases settled the launch: auto-upgrade on Windows, a permission prompt where Allow always covers every file operation in one click, Cmd+V paste on Mac terminals, requests routed only to US infrastructure, and sessions that survive a background sign-in refresh. Earlier versions are not covered here.