what changed, in plain words. the app checks for updates itself — nothing to do but say yes.
Fixed: declining a question from the AI — clicking "None of these — let's talk about it" in the side panel — could wipe out the reply that was already on screen, leaving a blank space where it had been. blu now saves whatever the AI had already written the moment you decline, instead of waiting on the AI to report back on its own, which it did not always do.
Fixed: a chat could get permanently stuck. If a local model got cut off partway through starting an action — most often because it hit its reply limit while still "thinking" — blu could leave the conversation broken in a way that failed instantly on every message after that, with no way back short of starting a new chat. blu now closes out an interrupted action cleanly when it happens, and heals a chat that is already stuck instead of refusing it forever.
Fixed: a model that spends most of a reply "thinking" without saying or doing anything now gets stopped, instead of silently burning its whole reply budget. This showed up most on local models doing heavier reasoning — one measured turn spent 89% of its allowed reply on thinking and produced a single word.
Local models now get a reply-length limit sized to what they can actually hold, rather than the same flat number for every model. blu also asks LM Studio directly for a model's real memory size instead of guessing at it.
GLM 5.3 is here. Z.AI's newest coding model is in blu's model menu, both in blu's own engine and under Claude Code. Your GLM Coding Plan also now appears as its own group in that menu — it did not before, so if you have a Coding Plan key saved, the three models it covers are simply there to pick.
GLM 5.3 shows a dash where a price would go, and that is deliberate. Z.AI has not published what it costs per token yet, and blu will not print a number it cannot stand behind. Its turns are marked as covered by your subscription rather than charged, and a conversation total that includes one says "at least" instead of claiming to be exact. When Z.AI publishes a price, blu picks it up on its own — no update needed.
Fixed: a model could be charged at an older model's price. If blu had no verified price for a model, it could quietly borrow the price of an earlier one whose name it started with — GLM 5.3 would have been priced as GLM 5, a different and cheaper model, and the figure would have looked measured. Anything blu cannot price now shows a dash instead. The same fault could put a cloud price on a model running free on your own Mac, and that is fixed with it.
Worth knowing if you are on a GLM Coding Plan: Z.AI now answers requests for GLM 5.2 and GLM 5.1 with GLM 5.3. blu still names and prices the model you picked, so on a plan those older GLM entries are labelled with a model that is no longer the one running. Picking GLM 5.3 directly avoids the confusion, and making blu record the model that actually ran is the next piece of work.
Carried forward from 1.2.0, for anyone updating from further back: the Claude Code engine can run other companies' models — GLM and DeepSeek publish Claude-compatible doors, and the model menu offers them under Claude Code with each price on the row. The key you already saved for a company works for its Claude Code entry too. Work a subscription absorbs is always shown as priced against the plan, never as if a card was charged per token.
The Claude Code engine can now run other companies' models. GLM and DeepSeek both publish Claude-compatible doors to their models, and blu's model menu now offers them under Claude Code — each with its price on the row. The key you already saved for a company works for its Claude Code entry too; settings says which key each row uses. The menu offers only the models each company documents for that door, so nothing in the list fails when you pick it.
Money got more honest with it. Work a subscription absorbs — Z.AI's Coding Plan, like your Claude plan — is now always shown as priced against the plan, never as if a card was charged per token. That is true on every engine, in the model menu, and in each turn's own cost line.
Fixed: switching chats could leave the changed-files panel showing the previous chat's files — the headline named one folder while the list underneath belonged to another, and it offered Undo on files the open chat never touched. The panel and its counter now always follow the chat you are looking at, including brand-new chats.
Fixed: the message box could grow a sideways scrollbar that squashed what you were typing into an unreadable sliver. Whatever you type or paste now always wraps inside the box — even a very long link — and the box grows downward like it should.
If you skipped 1.0: every reply now says which engine produced it and who paid, your own messages sit on the right in their own bubble, clicking a file the agent links opens it inside blu, folders without version history get one blu keeps for them (with an off switch in settings → your data), and your history can be encrypted with a recovery key that is the key itself — keep it somewhere you will still have it in a year. The full 1.0 notes are on the release-notes page.
blu 1.0. The numbers you see are numbers we can stand behind, and the ones we could not stand behind are gone.
Every reply now says which engine produced it and who paid, right next to the model's name — "Codex CLI · gpt-5.6-sol (subscription)" means your plan covered it, "(API key)" means it was charged to your key, and "(your Mac)" means it ran on your own machine and cost nothing. The tokens-per-second figure is gone from engines that never actually measured one: it was output divided by waiting time, which is not a speed. Where blu genuinely timed the writing, it still shows. A turn's cost now splits what you sent from what the model wrote back, because the two are priced very differently, and the conversation header counts "tokens processed" with an explanation on hover — most of that number is the same conversation being re-read each turn, which is how these models work, and now the label says so.
Your messages sit on the right side of the chat now, in their own tinted bubble, so scrolling back for what you asked no longer means re-reading everything blu did in between.
Clicking a file the agent linked opens it right in blu — rendered, readable, with a back button when one document links to another. Before, clicking copied the file's location and left you to go find it in Finder.
Folders without any version history get one, kept by blu. Point blu at an ordinary folder of documents, and the changed-files panel now lists what an agent created or edited there, shows exactly what changed, and can undo it — where before it sat empty. Nothing is added to your folder; the history lives with blu's own data, secret files like .env are never recorded, and there is a switch in settings → your data if you would rather blu kept no history at all.
Deleting files works the way you would expect. blu used to refuse every forceful delete outright, even inside your own project, whatever your trust setting said — now it asks you, and your trust setting decides. Deleting anything outside your folder, or the folder itself, is still refused, always.
Zooming the window with Cmd + no longer buries the message box, the settings costs page now shows the prices blu has always shipped for Anthropic, OpenAI, GLM, DeepSeek, Kimi and Qwen instead of claiming no rate was set, and the OpenCode engine has been set aside for now — it needs more work before it meets the bar, and chats that used it still open and read correctly.
If you are coming from 0.6.x: plans are approved on the plan itself, changed files appear while the agent is still working, and approving a plan genuinely starts the work.
If you are coming from anything earlier: questions from an agent wait for you rather than timing out, you can paste pictures into the message box, your chats group by folder, settings has five tabs, and your history can be encrypted with a recovery key that is the key itself — keep it somewhere you will still have it in a year.
You approve a plan on the plan itself. When Claude Code proposes one, the plan opens in the side panel with Approve and Deny at the bottom of it, and the free second opinion from a model on your own Mac sits above them. Until now a small approval box appeared in front of the plan, asking you to approve a document it was covering, with a button on it to go and read what you were approving.
And approving it now actually starts the work. Clicking Allow used to leave the chat in plan mode: the mode chip still said "plan", the agent stopped again for permission before the very first file the plan called for, and the next time blu started that chat it went back to planning instead of building. Approving now moves the chat to "auto" — or back to whichever mode you had picked for that folder — and the chip shows the mode the agent is genuinely running under rather than the one blu asked for. On a model that cannot run "auto", it moves to "acceptEdits" instead and says so on the chip.
Changed files appear while the agent is still working. The panel used to fill in only when a turn finished, so a run that spent five minutes rewriting files showed "Nothing changed yet" for all five — which is the wrong half of the job for a panel whose reason to exist is letting you stop a run that is doing something you did not ask for. Files now show up as they are written, and stopping a turn halfway reports what it already wrote instead of leaving you to find out later.
If you are coming from 0.5.x: questions from an agent wait for you rather than timing out after ten minutes, blu can notify you when it needs you, you can drag or paste a picture into the message box, a revised plan stays one plan with revisions, and the subagents panel opens on a list of every agent the chat has sent out.
If you are coming from 0.4.x: your chats group by the folder they were held in, and the side panels are colour-coded.
If you are coming from 0.3.x: settings is five tabs — models, defaults, costs, skills, your data — and your history can be encrypted from the last of them, with a recovery key that is the key to the database itself. Keep it somewhere you will still have it in a year. Losing it costs the conversations, not the app.
blu waits for you. When an agent stops to ask you something, that question now stays open until you answer it — however long you are away. Until now it gave up after ten minutes and let the agent decide for itself, and the answer it invented went into the plan as though you had agreed to it. Nothing is decided for you any more. If blu is closed while a question is waiting, the question is still there when you reopen that chat, and your answers go into the message box ready to send.
blu tells you when it needs you. A notification when an agent stops to ask something, and another when a turn finishes — only while you are looking at something else. macOS will ask your permission the first time there is something to say.
You can show blu a picture. Drag a screenshot into the message box, paste one, or use the paperclip, and the picture is part of the message. Dragging one in used to lose it entirely: macOS hands the file over rather than copying it, so the screenshot was not on your Desktop either. The paperclip put the file's location in the box instead of the picture, so the model was handed a line of text and told you nothing was wrong.
A plan stays one plan. Decline a plan, say what you want changed, and the rewrite is now the same document with your change in it — labelled "revision 2 of 2" rather than appearing as a second, separate plan with the same title. blu also keeps hold of the conversation the agent is having, so a restart no longer makes it start the plan again from scratch under a new name. That was why a plan you had shaped over several turns could come back rewritten, with a different title, having quietly lost the shape.
The subagents panel opens on a list. Every agent the chat has sent out, newest first, with the turn it ran in and the time. Click one to see what it did, and "open results" for its full report. Before this the panel opened straight into a single agent, and reaching any of the others meant scrolling back through the conversation to find where it started.
Fixed, and it undoes something 0.5.0 only half fixed: every plan said "no answer recorded" — including ones you had approved and ones you had declined out loud. The part of blu that reads your answer was looking in the wrong place and always found nothing, so the label had nothing to go on. 0.5.0 changed the wording for a replaced plan; this fixes the reason it was reaching for that wording at all.
Fixed: a chat with a long history was re-sent to the agent every time blu restarted, which cost tokens and told the agent this was a fresh conversation it was being caught up on. It now picks up the conversation it was already having.
If you are coming from 0.4.x: your chats group by the folder they were held in, the side panels are colour-coded, and you can approve a plan from the panel you are reading it in.
If you are coming from 0.3.x: settings is five tabs — models, defaults, costs, skills, your data — and your history can be encrypted from the last of them, with a recovery key that is the key to the database itself. Keep it somewhere you will still have it in a year. Losing it costs the conversations, not the app.
Your chats group by the folder they were held in. The sidebar gathers them under the project each one belongs to, alphabetically, so a folder stays where you last saw it instead of shuffling as you work. Inside a folder the newest chat is still on top. Headings fold away and blu remembers which you folded. The button beside "new chat" switches back to a plain list, newest first, for when you are hunting for something you cannot name — and it only appears once you have worked in more than one folder, because one heading over one list is just a heading. Chats saved before blu recorded a folder gather at the bottom under "no folder recorded".
The side panels tell each other apart at a glance. Changed files is green, subagents is black, a plan is indigo, and anything waiting on you is still the warm orange it was. They all wore the same black strip before, so two different things looked like the same thing.
You can approve a plan from the panel you are reading it in. The Approve and Deny buttons used to live only on the card in the conversation — the card you push aside to read the plan — so reading it and answering it were in two places and you had to go back.
While an agent is working, the popup shows what it is doing. Open a subagent's report mid-run and you now see the words as they arrive and the list of things it is looking at, instead of a tall empty box. The narrow panel beside your chat shows the opening few lines and a way through to the full report, which is where a report belongs: a wide table has never been readable in a 416-pixel column.
Fixed, and this one cost real work: a question blu asked while you were away could eat your answer. If nobody answered within ten minutes, blu let the agent carry on using its own judgment — but said nothing, so the question sat there still looking answerable. You could come back, answer all of it, press send, and have it go nowhere. The card now says what happened as soon as it happens, and if you answer it anyway your answers land in the message box so you can send them yourself rather than typing them a second time.
Fixed: the subagents panel would not open. It needed you to click an agent in the conversation first, so the button at the top right did nothing on its own — the plan and changed-files buttons beside it worked, which made it look like nothing was happening at all.
Fixed: you could not scroll that panel while agents were still running. The list of what an agent had done was squashed to fit the window instead of running past it, so there was genuinely nothing to scroll to. The same fault was in three other panels.
Fixed: wide tables ran off the side of the panel and there was no way to reach the columns that fell off. They now scroll sideways inside their own strip, in the popup and in the conversation, without the page moving.
Fixed: "Read the plan" did nothing. The plan was already open underneath the approval card sitting on top of it, so the button had nothing left to do. It now slides the card aside.
Fixed: a plan you declined and asked to be rewritten said "no answer recorded", as though you had ignored it. It says it was replaced by a newer plan.
Fixed: the +340 −96 count on the changed files panel was barely legible and always had been — dark green and dark red on a near-black strip. The numbers are plain now and the plus and minus signs carry the meaning. Per-file counts are still coloured, where the background is light enough for it to read.
Fixed: an update that began downloading at the moment a new version was published no longer fails partway. The download host keeps the previous release as well as the current one.
If you are coming from 0.3.x: settings is five tabs — models, defaults, costs, skills, your data — and your history can be encrypted from the last of them, with a recovery key that is the key to the database itself. Keep it somewhere you will still have it in a year. Losing it costs the conversations, not the app.
Type `/` in the message box and blu shows what that runner can actually do. Claude Code reports its own commands — `/compact`, `/cost`, `/context`, your own saved commands and skills — and the list narrows as you type. Codex gets its commands and any prompts you keep in `~/.codex/prompts`. On blu's own runner the list is your skills, and picking one drops its instructions straight into the box. The list is never invented: it is what that runner says it has, so it is right even when the runner changes underneath us. One command genuinely cannot work this way — `/login` needs a browser and a terminal window blu does not give it — which is why the next paragraph exists.
Sign in to Claude Code without leaving blu. If Claude Code is installed but not signed in, blu now says so plainly and gives you a Sign in button that opens the real sign-in page. Before, a chat would answer "Not logged in · Please run /login" in the model's own voice, as though the model had said it, and `/login` was the one thing that could not fix it. First run stops guessing too — it tells you whether Claude Code is signed in rather than hoping.
Your history can be encrypted, and you hold the only way back in. Settings → your data. Switch it on and blu gives you a recovery key: it is the key to the database itself, so keep it somewhere you will still have it in a year. Losing it costs the conversations, not the app — blu starts, sets the unreadable file aside, and carries on with a fresh one. Nothing is deleted behind your back.
Settings is five questions instead of one long page: models, defaults, costs, skills, your data. There is a setup checklist at the top that tells you what is still missing.
See what blu knows about you, and take it with you. Settings → your data shows where your conversations live on disk, how much space they take, and exports the lot.
Changes works on every runner now. Claude Code, Codex and OpenCode edit files inside their own processes, so blu compares the folder against a git snapshot to work out what actually moved — it used to show nothing at all unless you were on blu's own runner. In a folder that is not a git repository their edits are still invisible, which is the one limit.
Two speed numbers stopped sharing one label. "42 tok/s" is measured writing speed. "42 tok/s overall" divides the whole turn — including the thinking — by the clock, which on a reasoning model is a different quantity and a much lower number. Both were true and both printed identically before.
blu writes down what a subscription costs. A Claude or ChatGPT plan has no per-token price, so work run on one is priced separately from money actually charged to an API key, and the two are never added together.
Fixed, and it was a security problem: an agent working in one folder could read every conversation in your database, silently. It cannot now.
Fixed: a reply could vanish. If a turn finished while you were looking at another screen, blu sometimes refused to save it and the answer was gone when you came back.
Fixed: the review that double-checks an approval now only reads actual plans. It was running on every single command — including a one-line `grep` whose whole text was already on screen — and the paragraph it wrote underneath was noise on the exact screen where noise is most expensive. It also named the model that had done the reading incorrectly, and sometimes named none at all.
Fixed: blu no longer prints a speed it cannot stand behind. Where the writing window could not be measured it shows `—` instead of dividing by the whole clock, which on a turn full of tool calls reported roughly a third of the real rate.
Fixed: the note that says a chat is waiting on you led with the chat's title, which is your own first sentence, so it read as three lines of your prompt with the actual news stranded on the end.
If you are coming from 0.3.0 or earlier: API keys are saved now — a key pasted into any version before 0.3.1 reported success and was silently discarded, which is why models you had paid for never appeared. Keys live in your Mac's own keychain and nowhere else. Kimi reaches the international service rather than the China one, the GLM Coding Plan is its own separate connection, and OpenCode greys out models it has no credentials for instead of going quiet for thirty seconds.
Kimi works. blu was asking api.moonshot.cn, Moonshot's China service, which refuses a key issued on the international one however many times you paste it. It now asks api.moonshot.ai — and if you already had Kimi set up, blu corrects the address for you when it opens. Five models, each with its published price: K3, K2.7 Code, K2.7 Code HighSpeed, K2.6 and K2.5.
The GLM Coding Plan is its own connection now. A Coding Plan subscription and a pay-as-you-go GLM key are different products on different addresses, and each refuses the other's key — so a Coding Plan key on the old setting reached nothing at all. Switch on "GLM Coding Plan (Z.AI)" in Settings and paste it there instead. Use one or the other, not both: they serve the same models, and blu cannot yet tell two of them apart.
OpenCode tells you what it cannot run, before you pick it. Models it has no credentials for are now greyed out in the picker with the reason, instead of accepting your message and going quiet for thirty seconds. OpenCode keeps its own credentials, separate from blu's — the note names the model and points you at `opencode auth login`.
Qwen is available, switched off until you add a key. One model is priced: Qwen3.8 Max. The rest of Qwen's lineup changes price depending on how long the conversation has got, which blu cannot yet describe honestly, so those read `—` rather than a number that would be right at the start of a chat and wrong by the end.
Shorter model lists, with nothing missing. Asking a provider what it serves returns its whole catalogue, including every dated re-release of the same model — one provider offered 79. Where the plain model is also on offer, the dated copies are now folded into it. Anything you could only reach by its dated name is still there.
Fixed: switching one connection off blanked the prices and model counts on all the others until you pressed Refresh.
Fixed: the conversation's name and its running total, at the top of the chat, were too small to read at a glance.
If you are coming from 0.2.1, the big one is in 0.3.1: API keys are saved now. They never were — a key pasted into any earlier version reported success and was silently discarded, which is why models you had paid for never appeared. Your keys live in your Mac's own keychain, encrypted, and nowhere else. Also from 0.3.1: blu's own agent can hand work to helper agents on models you choose, Claude Code and Codex keep their own controls under their own names, and "Allow always" now remembers exactly what you agreed to rather than roughly — so it may ask once more about something you had already allowed.
API keys work. If you ever pasted a key into blu, saved it, and found it gone when you came back — that was blu, not you. Keys were never reaching your Mac's keychain at all, in any version we have released. They do now. Please paste your keys in once more: there was never anything stored, so there is nothing to recover.
Your keys go into your Mac's own keychain — the encrypted store macOS already uses for your other passwords, locked with your login password — and nowhere else. No copy on disk, in any version of blu, including the ones we build for ourselves.
Kimi accepts your key now. blu was asking api.moonshot.cn, which is Moonshot's China service and refuses keys issued on the international one, however many times you paste it. It now asks api.moonshot.ai. Kimi K2.7 Code and K2.6 join K3 in the list.
Seven GLM models instead of one — GLM 5.2, 5.1, 5 Turbo, 5, 4.7, 4.7 FlashX and 4.6 — each with its published price, so a turn on any of them is costed from real rates rather than left blank.
Fixed: the model list offered things that cannot hold a conversation. Asking a provider what it serves returns everything the account can reach, including models that only make images, transcribe speech, or turn text into numbers for search. One provider alone listed over a hundred. They are gone from the picker — including on your own Mac, where an embedding model had been sitting in the list of local models all along.
Fixed: a Groq key could be saved and then never recognised.
If you are coming from 0.2.1, three things changed in 0.3.0 that you have not seen yet. blu's own agent can hand parts of a job to helper agents, each on a model you choose. Claude Code and Codex now keep their own controls, under their own names, instead of blu's. And "Allow always" remembers exactly what you agreed to rather than roughly — so you may be asked once more about something you had already allowed. The connectors panel was removed; it could report "connected" while the AI was never able to use it.
Helper agents. blu's own agent can now hand parts of a job to several helpers at once, and each helper runs on a model you choose — a cheap one to read and fetch, an expensive one to think. You watch them work, and every helper is priced at its own rate rather than the main model's. If the model you are chatting with is a small one running on your Mac, blu checks it has room to read the answers back before it spawns anything, and tells you plainly when it does not instead of quietly cutting them short. This is blu's own agent only; Claude Code and Codex cannot do it.
Chats keep working while you look at something else. Switch to another chat, start a new one, rename one, or open Settings — the reply keeps coming, and the sidebar shows you which chats are still working. If a chat stops to ask your permission, blu now tells you wherever you are and takes you to it. It will never ask you to approve something from a screen that cannot show you what you are approving.
Nothing is lost if blu closes in the middle of a reply. Your question is saved before the model is ever called, chats are named the moment they start instead of after the first answer, and blu asks before closing while work is running. A chat that was interrupted says so, rather than opening empty.
Claude Code and Codex now keep their own controls. If you use either one, the setting beside the model is theirs, not blu's — Claude Code's own permission modes and Codex's own sandbox settings, under their own names. blu passes your choice through and stops overriding it, so what you get is the tool you already know, in blu's window. blu also stopped putting its own rules in front of Claude Code: when Claude Code decides something needs your approval, that question now comes to you directly, and when it doesn't, blu no longer interrupts.
One second opinion, on a model you picked. blu could have another model check a risky action before you approved it, but the model doing the checking was chosen for you and there was no setting to change it. There is now one review, it names the model doing it, and you can change that model on the card itself or switch the whole thing off. It is also shown the entire thing it is judging — it used to be handed the first couple of hundred characters and asked for a verdict.
Skills from a project are held until you have read them. A repository you clone can carry written instructions for the AI, and one of those could quietly replace a skill of your own with the same name. blu now shows you the file first and asks. Anything already in your own skills folder was let through once, so nothing you installed yourself has changed.
"Allow always" now remembers exactly what you agreed to. It used to remember roughly — approving a read of one file could cover the .env sitting next to it, and approving one python command could cover every later one. It no longer does. You may be asked once more about things you had already allowed; that is the correction landing.
The connectors panel has been removed. It could add a server, store its password and report "connected" while the AI was never actually able to call any of its tools. Rather than leave a switch with nothing behind it, the panel is gone until the part behind it works.
Fixed: opening Settings ended every chat that was working.
Fixed: several numbers were wrong. Turns priced against the wrong rate, five Claude models with no price at all, a conversation total that left out real money, and writing speed measured against the time a model spent waiting for a command rather than the time it spent writing.
Fixed: the model list. It offered models blu had no key for and models the runner could not reach, and it reported a key as rejected when the key was fine. Saving a key now takes effect immediately, and the list names the company each model comes from, newest first.
Fixed: menus that opened off the bottom of the screen, changes shown as two walls of text instead of a diff, an approval that asked you to accept a plan you could not read, and a chat saved the instant the window opened whether you typed anything or not. blu also says which version it is now, under Settings.
Also fixed, all of it security: text blu reads on your behalf — from a file, a web page or a tool — is now marked so the AI treats it as information rather than as instructions from you. A file could be swapped for a shortcut in the instant between blu checking it and opening it. A link could lead the agent out of the folder you had locked, or to an address only your own Mac can reach. On the OpenCode runner, blu was answering yes on your behalf to every edit and command — it now asks.
Plainer language throughout. Messages said things like "Harness error" and "exited exit status: 1". They now say what happened and what to do about it.
A test of the update path itself. If you are reading this inside blu, the thing it was testing worked.
Fixed: an update could fail if it was installed in the first minute or two after a new version went up. blu would offer the new version, download the previous one from a cache, and refuse it because the two did not match. Each release now has its own address, so there is nothing stale to serve.
Nothing else has changed from 0.2.0.
The first public release. blu runs every AI model — the free ones on your Mac and the paid ones you rent — from one window, and lets you change your mind mid-conversation.
Four ways to work, and you can switch between them in the middle of a chat: blu's own agent, the Claude Code CLI, Codex on a ChatGPT plan, or OpenCode. The conversation goes with you — whatever you switch to can see what was said before it arrived.
Every turn says what it cost, and the conversation's totals sit at the top of the chat. Models running on your Mac say "free", because that zero is real. Work covered by a subscription is counted separately from money that actually left an account, and the two are never added together. Where blu does not know a price, it says so instead of showing you a zero.
Fixed: blu can now find the programs installed on your Mac. On earlier builds a downloaded copy could not run node, pnpm or Homebrew tools, and the Claude Code runner would not start at all.
Fixed: changing model or runner used to lose the conversation on every runner except blu's own.
Fixed: several numbers on screen were wrong — a free turn shown as $0.00, tokens counted as words, and writing speed measured against time the model spent waiting for a command rather than time it spent writing.
This release is chat and settings. Jobs and compare are not in it — they are being rebuilt properly and will arrive as downloads.