Changelog

Added

• Antenna between two computers on your network: on one you choose «Invite» and get a 4-digit PIN; on the other you choose «Join» and type it.
• Vault: an AI can leave you empty variables to fill in, and with the window open what an AI saves shows up by itself.
• Factory costumes rewritten, shorter and clearer. When you update, your agents get the new ones and a copy of the old ones is kept; what you edited yourself is left alone.
• The new application: CleverCoder was rebuilt from top to bottom and replaces the previous version. When you update, your projects, agents, license, vault, settings and backups stay where they were; you don't have to do anything.
• Every open project has its own button on the taskbar.
• Rebuilt input bar at the foot of the terminal: microphone, speaker, send by voice, and you can drag images and files into the box.
• Voice: Kokoro downloads with one click from Audio output (about 115 MB), besides Piper. Audio input and Audio output have their own window, with a single voice choice per agent or general, and dictation types while you speak.
• Panel: lists, images and videos inside the cards, and drawings you see live.
• Backups: history, a single "Restore" that detects what each copy holds, export to a .ccbak file and a countdown to the next copy to the flash drive.
• Agenda that creates, edits and runs on its own; Run button in the panel; cost and plan per model in Models.
• Its own font, icons and emoji, the same on any computer; add an MCP by pasting the "claude mcp add" line.
• Gemini as an agent: paste its key once in Settings › Models and the agent works in its own folder, with access to the code and to its tools.

Changed

• «Restore factory» on a costume keeps a copy of the previous text.
• Piumy and Verdesk no longer load into every agent of a new project: they reach the agent you assign them to.
• The X of an agent open in another CleverCoder window tells you to close it from that window.
• The download is bigger: about 165 MB (before, about 120 MB).
• Uninstalling also removes "Open at Windows startup" and the application's web page cache.
• With the speaker off, card audio is prepared slowly, without loading the PC; pressing Play speeds it up.
• The drawing viewer tells you when it takes long to open, instead of staying blank.
• The Agenda no longer loses reminders when the app reopens, shows overdue and failed ones with "Retry", and tells you whether routines are running.
• The photo snip freezes the screen when it opens and saves exactly what you picked.
• Changing the language in Settings offers "Restart now".
• With several windows open, skins are no longer lost.
• Backups says "backup pending", without a red alert.

Fixed

• What an agent learns, or a change to its costume, now reaches it when its conversation is compacted or cleared.
• The welcome no longer closes with an error on step 2, "Extras".
• If an agent's program closes by itself, its tab closes and the agent is left asleep, ready to reopen without duplicating.
• Codex starts without getting stuck on the update question or on the one about trusting its hooks.
• A split emoji typed into the terminal no longer closes the application.
• Panel cards no longer vanish and come back, and a card's audio no longer plays twice.

Added

• New project: a simple window, with just what it takes to create one.
• Backup: option "Protect with the vault password", and backups keep the project's Claude memory.
• Vault: "Copy" button on each value.
• Party: if an agent ran out of quota, the board says until what time.
• Free license: a single one, with the same features as the paid one, free if it is not for paid work or for a company (personal use, study and non-profit organizations). For paid or company work, the paid license.
• You request it from the program ("Use for free") or at clevercoder.app/gratis, with your email. It lasts 36 weeks; the program warns you two weeks before it expires and you renew it with one click.
• With the free license, the AI knows it from startup and reminds you once if the work looks commercial, without blocking.
• Pro codes: redeemed only once, in the program, and they give 3 months of the paid license, with commercial use. Each code expires 36 months after it was created.
• Panel: while a card that is out of view is playing, a speaker appears above the scroll-down arrow; tapping it switches the panel to All and centers that card. It goes away when the card comes into view or the audio ends.
• Listen mode: cards play even when their panel is not in view, in queue order. It turns off in the speaker settings.
• Two-step antenna: with "Request connection approval" in Options, whoever receives the connection sees a notice with "Accept connection" or "Reject". With no answer in 60 seconds, it is rejected. An agent can also accept a pending connection when you ask it to.
• Antenna: the antenna menu shows live connections at the top; inside Connect, the agents of other projects, with their status (On, Wakeable, No connection).
• Vault: each key is global (all projects) or local (its project only), with a filter. A key can hold several variables, for example username and password, each with its eye to hide it.
• Models: adding an API for Claude Code with its address and its variable; "✓ Test" says what failed. Adding an OpenCode API writes its configuration after asking for confirmation and keeps a backup copy. "Remove this API" removes it.
• /clever-reload: restarts an agent with the same conversation, so it picks up new MCP servers, skills and plugins.
• Costume catalog: the 146 Agency Agents always appear in their group, in any project. Each one is installed when you create an agent with it.
• Create agent and the agent sheet have a single costume selector, with a search box. Your own costumes and CleverCoder's come first, then Agency Agents, with a heading for each type and each costume's icon.
• Pool: the AI installs a tool from the published catalog by itself and tells you which ones can be installed.
• Pool: the tool list is exported to a file and imported on another computer, with the Export and Import buttons. Keys never go into the file, and anything imported that needs keys comes in turned off so you can review it.
• At startup, if there is a new version, a dot with «1» appears on the Update button and a card shows up in the panel, once per version, without slowing down startup.
• Online activation: a paid license or a Pro code is registered on each computer, up to 2, and opens for up to 7 days without a connection. If you reach the limit, "Free my computers" emails you a link. The free license and the trial do not need a connection.
• Vault: search by name, and the list in two groups, local keys first and then global ones.
• Backup: backups include the vault, both local and global, encrypted as it is, also on the USB drive. "Restore vault" brings it back without overwriting a vault that already exists.
• Panel: right-click a link on a card to copy its address without opening it.
• Chocolatey: CleverCoder is also published as the "clevercoder" package. It can be installed with choco once Chocolatey finishes reviewing it.

Changed

• The 2-month trial no longer exists: it is now 3 days without registering. Every launch during those days shows "Register later · N days left". After the 3 days, you need a free license, a paid one, or a Pro code.
• Antenna: Connect and Disconnect no longer open a window; the result appears as a panel card. "Connected" is shown only when the other side confirms it, and if the other version is older, it tells you to update it.
• Connect no longer shows a success message: the antenna counter says it. The counter counts the established connections and refreshes itself about every 15 seconds.
• Antenna cards say who sends the message and whether it goes out or comes in. Outgoing ones have no Reply button.
• With the split screen, the panel counts as in view: what you type in the terminal bar enters as a panel message.
• Messages no longer carry an automatic notice at the start: the panel skill is on for all agents.
• "Startup lines" moves from Settings to Models, next to the APIs, and the agent sheet has "Edit lines".
• "Import image" becomes "Import file": an image is copied to the imported-images folder as before; any other file is not copied, and its path is written in the agent's bar or in the card in view.
• Adding the same API again adds its new models; with a different address or variable, it asks you to remove it first.
• Keys managed by CleverCoder are marked "Managed by CleverCoder" and are always global.
• Reports (the report panel and the error window) go to the CleverCoder server instead of the Google form, which had not received anything since April. The server sends them by email to the team.
• The costume editor opens much faster with a large catalog, holds off requesting the Agency Agents costumes until you pick one, and shows "Loading…" while it arrives.
• Drawing viewer: the wheel scrolls the drawing up and down, Ctrl plus the wheel zooms in and out, and the drawing opens fitted to the window width.
• A disabled button now looks disabled across the whole program, with readable text in all ten themes.
• If an agent is already open in another CleverCoder window, opening it again brings that window to the front; if Windows does not allow it, it flashes on the taskbar.
• Three texts that said different things in each language now say the same.
• The installer in silent mode (--silent) installs without opening CleverCoder, and does not downgrade if the installed version is newer.
• The license lock window says what is happening: the title bar shows the same as the heading, and when the license is already on 2 computers the main button is "Free my computers".
• Backup: the "Additional folders" option is removed; it never backed anything up.

Fixed

• Per-step audios on a card no longer stay on "Generating audio".
• With the PC suspended or hibernated, stuck-message warnings no longer count that time, and a message that did not get in is resent only once.
• The Backups panel no longer erases the settings an agent left, a replaced image on a card refreshes, and nine messages get their line breaks back.
• A message between agents that could not be delivered for about 9 minutes was lost silently; now it keeps retrying and, if it still fails, it is logged.
• A card's audio stops at once if the agent edits, withdraws or deletes it while it plays, including in the listen-mode queue.
• An agent's reply that quotes one of your cards shows its audio player, or the three dots while the audio is generated, or the error note. Before, it showed none of them.
• The reply to a card draws its block with the Copy button.
• The speaker turns off if the tab that owns the playing card is closed.
• An OpenCode terminal counts as ready once its prompt box is drawn: the context of the initial card is no longer lost.
• When closing, reloading or shutting down a terminal, the message that was arriving is no longer pasted over.
• The antenna joins two CleverCoder versions from the same installation.
• The antenna menu looks up connections in the background, off the window thread.
• "Disconnect" does not ask for the vault password when there are no saved connections.
• The vault and agent sheet texts can be read in all ten themes.
• A terminal that started with its tab hidden took a long time to start until someone looked at it (45 seconds against 1, measured). Now it starts on its own, even if nobody looks at it.
• Antenna: "Connected" in the menu reads well in all ten themes.
• The orange dot that warns a .cui drawing is marked to update turns on again.
• "Delete the key" says what it does: it deletes this computer's key, which still counts toward the 2-device limit. To free it up, there is "Free my devices".
• Double-clicking a project no longer pastes the old wizard's text into a loose terminal.
• «Check now» no longer gets stuck on «Checking for updates…»: it shows that it is checking and, when done, says whether you are up to date, whether there is a new version, or why it could not check.
• Card voice no longer hangs when the computer is busy compiling.
• The same agent can no longer be opened twice from two windows, and the agent list no longer says the terminal opened when it did not.
• Pool: the top buttons no longer wrap in narrow windows and the reds follow the theme.
• If the computer shuts down right while the vault is saving, the key being saved is no longer lost.

Security

• An isolated agent no longer uses the vault: its vault tools reject it, and an agent cannot see the local keys of another project.
• Pool: an isolated agent cannot use the Pool tools, and an installation that asks to run a command is refused before it starts.

Added

• Split screen: terminal and panel side by side, with a draggable divider and a button to split the screen. A single bar divides the width between them for all agents, and the terminal reaches down to its typing box. If the window is too small to split, the button turns off.
• The tools from the terminal footer (code explorer, current drawing, notes, icon picker, antenna, eyedropper, color wheel, ruler, rectangle, capture and import image) are also in a vertical bar on the right edge of the panel, in a single column, each with its state in view. The pickers insert what you chose where the cursor was: in the card in view or in the agent's typing bar.
• In split screen, open notes cover only the panel, and the footer button is a sheet of paper with the number of pending notes.
• A star on every panel card and a star filter. The mark is saved and also covers the party chat.
• An answered question folds into a single line, "You answered: X · Change", and pressing the chosen option of a question or form again deselects it. Questions always offer "Other…" to type the answer.
• Arrows to reach the agent tabs that don't fit in the panel's strip.
• The panel search finds text regardless of capitals and accents, and also by the name of the agent who wrote the card.
• The card reply box has a grip bar to change its height and grows by itself as you type, dictate or paste.
• The Open button of an external link opens it in the default browser.
• While dictation transcribes, the typing bar shows three beating dots in a fixed spot inside the box. If the transcription fails, the error is shown.
• Dictation with an NVIDIA graphics card: an optional CUDA engine that downloads separately and survives updates. If the card doesn't start, it goes back to the processor and says so.
• Redone agent sheet: name, large logo, Change icon and a button that says Open or Sleep. Each agent has its own startup arguments, with the "This is how it starts" line in view, and its own tabs for Auto-continue and Hard rules.
• Initial prompt: the session context bar becomes a simple card with three status dots and the latest context in view, with its creation and delivery time.
• The "Waiting for the terminal to be ready" bar shows one segment per condition (the program started, the screen is up to date, the terminal is quiet) with the name of the agent's program. The context goes in when all three are met.
• New project: what you see when creating it is a single line, the project opens with the panel in split screen, and the wizard lets you customize the leader (icon, startup arguments and the "This is how it starts" line) and pick it from all the installed models.
• A project without a leader says so in the agent list and, when you open an agent, recommends creating one first (Create now or Continue without a leader). Create agent offers the leader role when the project has none.
• The Costumes tab lets you create, edit and delete costumes, grouped into Personal, CleverCoder and Agency Agents, and has a "New costume" button. An agent's icon is changed by tapping its logo.
• Models: the "Add a model by hand" button (program, display name and id), which sync doesn't delete; Sonnet 5.5 built in; each API with its models underneath, in a collapsible group; and favorites first, with a star, in their program's list. The agent sheet, Create agent and the new-project wizard offer the same list: the one of the installed programs.
• Tools are shown in three boxes, Essentials, Complementary and Third-party, in the Pool, the agent sheet and the role editor. What is on glows and the Essentials have a lock. They are turned on and off at three levels: the Pool for everyone, the role for everyone with that job and "Customize this agent" for exceptions, with the option to go back to the role. Turning on an MCP or a plugin shows the notice to restart the agent.
• The first-launch installation is arranged in seven collapsible packages, with Cancel, Install recommended (or the selection) and Install all. Each tool shows its icon and its state.
• Two status bars to choose from in Tools: the simple one and one from a third party, audited, with tokens and 5-hour and 7-day usage, which only connects to Anthropic. Installing one replaces the other.
• The routemap shows each contract's progress next to its number, for example "3/4 Done", and the "to confirm" state of contracts is shown in magenta.
• Any agent can read the text of another agent's terminal, without permission; the image only if it asks for it.
• The antenna says which agent and which project a message was delivered to, the card of a message that arrives through it says who sends it ("cAPI from agent@project (not verified)") and, when you hover over it, each number is explained: connections open now and messages received. Seeing 0 connections is normal.
• Buy CleverCoder: licenses bought at clevercoder.app arrive signed at your email, and the program verifies them offline.
• If you are already on the free trial, you paste the license you bought in Info, License, with the «I have a license» button, and Buy takes you to the store.
• Global Pool: your tools (skills, MCP servers and plugins) in one place for all your projects, with three states per tool: Default, On demand and Off. It builds itself the first time you open this version and leaves a card explaining what changed. Turn it off in Settings, STARTUP section, without losing anything.
• Antenna: the 📡 button opens a menu to turn it on, connect and see its options. Connect links two agents in one step, with one secret per direction kept in the vault, and the conversation shows in both panels.
• Agent sheet: the «Add my Claude Code tools» checkbox lets that agent also use your connectors, such as Gmail, and your MCP servers. It comes off.
• Agent sheet: the model is chosen first by program and then by model, with search and favorites, and each Claude Code agent has its effort level.
• Panel: links can be clicked wherever they appear, a marked block gets its Copy button, a very long card shows folded, and each step of a card can have its own audio.
• Vault: the master password can be any password of 6 characters or more, with a link to change it and an eye to show what you typed. The unlock notice says who asks for it and why.
• The drawing viewer turns pages when a drawing has several, also with the arrow keys.
• Each message you write to an agent shows whether it arrived, is on its way, or did not arrive. If it did not arrive, a Retry button sends it again; it never retries on its own.
• Info shows your registered email and the days left in the trial, and lets you delete the key with confirmation.
• Each agent tab shows the tokens it has used, and the project total adds them all. The number only goes up: deleting a conversation does not lower it.
• The Models editor shows three real states: available, discontinued with its date, or not checked. Before, every model was marked as available, and 46 of the 103 NVIDIA models were already discontinued.
• Create agent returns to the list with the new agent marked, at the top of its group.

Changed

• The antenna listens only on this machine by default. Those that had been left open to the network are closed once when you open this version; if you needed one open to the network, open it again.
• The panel filters are CSS only: hiding no longer changes anything on the card. "Selected only" follows the agent you picked by hand and isn't released when you send.
• The single panel reads the same whoever the owner is: broadcasts, always, and the conversations of all live agents. Switching tabs no longer hides chat rows.
• The panel search queries the database off the interface thread: the freeze when pressing Enter drops from 130–540 ms to about 30.
• Messages between agents wait for the AI's end-of-turn signal, and if there is typed text not yet sent in an agent's terminal, they wait until it is sent or deleted (up to 30 minutes; a stray key, 60 seconds).
• "Send now" sends the context at once. It only doesn't paste it if the agent's program isn't running: then it warns in red and keeps waiting, because pasting it would go to PowerShell.
• The notices for a stuck queue, a stuck clear and a stuck startup arrive only when something changes, spaced further apart each time, and several at once go in a single notice with the list.
• Form cards are heard in full, with the title and each question with its numbered options, and voice speed is the standard multiplier: x2 is twice as fast.
• Skills have a single source. An automatic migration moves the loose skills at the project root to a backup and leaves a RESTAURAR.txt and a notice; a shortcut is no longer written for every skill, and each skill appears only once in the / menu.
• Nobody becomes leader by themselves: the automatic promotion of the oldest agent was removed. A new or added project goes first in the project list.
• Saving or restoring a costume tells the truth: whoever gets it from now on receives it, and those who already have it keep their copy. The "Remember for this costume" button is gone: the checkboxes of the context card are remembered by themselves for each agent.
• Deleting an agent closes its terminal first, and the delete window says what really happens. The help texts of the footer and the bar use no jargon.
• The agent's hard rule is now called Hard rules and its help says when it really applies: at startup and after clearing or compacting the context.
• Codex agents start with the MCP servers of their home, and Codex and Kimi receive the skills. If a server is left out, the status bar says so in red.
• Clever-release builds a publishing script of the project's own, with a fixed destination, and separates building from publishing. The releases button sends /clever-release.
• Panel cards appear at once and no longer wait for their audio.
• The antenna belongs to the agent: it comes back on when you reopen it and no longer leaves a new record for each terminal.
• Split screen applies to all agents and survives a restart.
• CleverCoder no longer sends messages to Claude to read the usage limits: it takes them from what Claude Code already reports.
• Automatic backup uses much less processor and space: it leaves out folders that can be rebuilt and goes from 7.5 GB to just over 1 GB.
• CleverCoder skills are 35% shorter and up to date with the program.
• Clicking an agent's tab only changes who you write to: the panel keeps showing all chats, and the filters do the rest.
• With a filter on, the panel brings only the messages that filter shows. Before, it worked with 400 and you saw between 70 and 117.
• The agent's tab shows its turn: the border and the small bar stay green while it works, and drop to zero the moment it ends.
• A change to the auto-continue interval applies live, without reopening the terminal.
• Opening a hidden terminal is faster: retained output is processed at about 300 ms per MB, down from 525.
• The installer closes the leftovers of the previous program and updates without asking. If another program holds the old folder, it names it and changes nothing.
• The repository and downloads move to github.com/clevercat64; old links redirect.
• The demo mode, the Terminal AI row in the agents panel, and the button that turned a loose terminal into a leader are removed. The model and bypass are chosen in Create agent.
• Agents no longer ask you with terminal menus: every question reaches you through the panel.
• The Project space tab is now called Global memory.

Fixed

• Messages between agents are no longer typed into PowerShell before Claude starts, where they ran as commands and broke the startup. Those from the previous session are not pasted: a notice arrives with how many there are.
• A message between agents could wait more than 20 minutes after a startup or a /clear even if the agent was idle, and moving the mouse or changing focus over a terminal counted as typing. Now the queue holds only while the /clear or the startup lasts, and only the keyboard, pasting and the panel count.
• A message from another agent no longer goes in or sends what you are typing in the terminal.
• The /clear ritual runs in strict order, with a single gate: while it runs nothing enters the terminal, no timeout writes the /clear with a turn open, and it no longer sends Enters to an empty prompt or waits 10 minutes for an end of turn that doesn't exist. The user's typing bar is never held by the clear.
• The bell and auto-continue sound once per end of turn, without repeating every ~15 seconds or going off twice.
• A turn switched on by an old send after the AI finished no longer holds the message queue, and asking for a /clear while another is in progress no longer writes it twice.
• Messages pasted into a terminal with a hook are sent again a few seconds after silence, and an agent that comes back doesn't start with all the old broadcasts unread.
• After a /clear, old party messages are no longer pasted again with a false "did not go in", and a message delivered and not archived at startup goes back to the queue.
• The "Key detected" window no longer shows up again on normal startups.
• Split screen: the panel no longer covers the terminal, no black strip is left over the panel, the typing bar no longer crosses both halves with its buttons under the panel, and the terminal toolbar stays only over the terminal.
• The audio of long cards no longer fails on time, a card's audio is no longer requested again every 14 seconds while nobody listens, it plays again after using a filter, and "heard" is marked only with real playback.
• An answered question shows the chosen option (before it looked open and was answered two or three times), an unanswered one is no longer hidden by the search, and an empty questions filter says how many are pending and why they aren't visible.
• The party chat in the panel no longer hides messages painted as deleted, when you change filter the panel shows the latest card, and the panel document no longer has a second scroll bar.
• The image, scissors, icon and CUI pickers paste into the box where you were typing, not into the terminal in view, and closing the ruler, the rectangle, the color wheel or the picker with another tab active no longer steals the background tab's focus.
• The panel's lit controls reach 3:1 contrast in the light themes and in Cabalery, the accent of buttons and the active tabs of Info and Account are readable in all themes, the Create new agent button is readable in all themes and the card logo no longer shows black in the dark themes.
• The search, load-more and filter icons of the panel show their author's name, and changing an agent's icon shows up instantly in the tabs, the list, the panel and the typing bar.
• Agent sheet: it can be scrolled with a short window, a sleeping agent shows Open and asleep as in the list, and the logo background color and inherit-from-role show as applied at once. Plan mode and skip permissions can no longer be checked at the same time.
• An agent that opens or sleeps in another CleverCoder window shows up again in the sheet and in the list, and the Open button in Party no longer shows the "ObjectDisposedException: EditorWebView" error.
• Dictation with the large model uses the server that is already loaded instead of loading another copy of the model, the dictation notices no longer pile up in the typing bar or give it a scroll bar, and the dictation model download and the graphics card engine download don't run at the same time.
• Models: a model added by hand is no longer lost if the models file or the startup-lines file is taken or damaged, and the agent sheet also brings the favorites and the full drop-down.
• Uninstalling a skill also removes it from the agents' homes, so it no longer comes back when you reopen the project, and a built-in skill that was on in an agent can now be turned off.
• Project without a leader or without any agent: Create agent can now create the first leader. The role selector shows each role's icon, and changing the model or the costume in the sheet updates the agent list at once.
• If a crash report could not be sent, the status bar says so in plain words. Changing the model or color of an agent without a costume no longer replaces its rules, and the costume selector offers the same costumes as the Costumes tab.
• The name of whoever sends a message through the antenna can no longer carry invisible characters that flip the text, and the title cut no longer splits an emoji.
• A plugin installed in Claude Code no longer turns itself on for every agent: security-guidance loaded the machine with one program per action.
• The license email can be edited in the Key detected window and in Info.
• An agent with skip permissions also starts that way when reopened from its sheet.
• With the vault closed, turning on the antenna no longer creates a new credential without telling you.
• Moving a key to the vault no longer leaves a plain copy if the vault fails.
• Two projects backing up to the same disk no longer collide: they take turns.
• A message from the panel to an agent with its turn open no longer shows as "did not arrive", and Retry no longer duplicates it.
• The panel no longer freezes the window when loading more cards, no longer flickers or shrinks when it opens, and no longer shows old messages after several filter changes.
• A card that asked for audio and stayed silent shows the audio generation error, and a voice order to another agent no longer leaves the window silent for five minutes.
• Going from Party, Pool or Contracts to Info or Settings no longer leaves the terminals covered, and a terminal whose viewer crashed shows again.
• After a browser crash, no extra input bar appears, the response boxes of forms no longer have a double border, and the floating agent column moves down with the context bar.
• Notifications can be read in all ten themes, with the accent and the name at sufficient contrast.
• The thinking indicator no longer lights up when the program only answers a permission prompt.
• Models reloads the catalog from the internet, which never loaded before.
• Deleting the key no longer brings it back through revalidation on opening, and the memory tab and the instructions mirror no longer lose edits.
• Roles that are not in the database show their factory icon in Create agent and in the catalog.

Security

• Vault with a master password: if you create it, CleverCoder does not ask for it at startup, only the first time something uses the vault. Once open, it stays open until you close the program. If you cancel, that request fails with a notice and the program goes on. Without a password, everything stays as before.
• The key of a Models API is shown only by its last 4 characters, and the full key is no longer left in the page. A model's id goes through a list of allowed characters before entering a startup command.
• The antenna listens only on this machine by default.
• Codex agents start with no secret on the startup line.
• Publishing a version no longer requires pasting credentials: the project's publishing script runs with the vault's credentials only in its own environment, you approve it by hash in a program window, and the output is suppressed if the secret appears.
• An agent can no longer turn on skip permissions for another agent, and a vault approval cannot be forged with a file in the project.
• The antenna only sends its pin to a CleverCoder of this same installation.
• The DeepSeek and Kimi keys are delivered from the vault, and only to the OpenCode terminal. After moving a key to the vault, the button offers to delete the plaintext backup, only if the vault holds the same key.

Added

• A free 2-month trial, no card needed, with a button on the activation screen. It keeps running after you close the program, and if the license file is lost it resumes with the same expiry date.
• The activation screen asks for your email, shows your machine's identifier so you can give it to support, says what data leaves the machine, and offers the button that fits each error: renew, contact support, or retry.
• The first-run setup screen installs Claude Code with its official installer, checks that it's ready without restarting the program, and offers the essential and secondary tools, with Cancel and Install essentials buttons. Node is no longer an essential.
• From the tools area you can install Codex, Gemini, Ollama, Kimi, and OpenCode. Kimi is now installed as its own program; the entry that used to carry that name, which was really Claude Code pointing at a Kimi server, is now called what it is. The area also shows each engine's environment variables, read-only and with secrets hidden, and the icon packs: five downloadable, five included, and three from Windows.
• "Set status line" installs the Claude Code status bar with one click, without using AI and without overwriting one you already have: it backs yours up, and Uninstall puts it back.
• New screens to create an agent and to browse the role catalog, with a button in the top bar. A new project can already assemble a full team with the roles the product ships with.
• Each agent shows how much context it has used, as a percentage, and warns the agent and the leader when it crosses 70% and 80%. Session context moves to a bar above the terminal, with a countdown, instead of a floating window that blocked everything else.
• The dashboard and the agent list warn when someone has been open for more than 30 minutes with no assigned work, or when a session is still alive but stopped producing output.
• Auto-continue gains Alarm mode, which fires at the time you set, and when it reaches the last text in the queue it goes back to the first.
• Mega-Plan is renamed clever-loop on screen, in the skill, and in the command; the old command still works.
• The program checks for updates again every 6 hours while it's open, without interrupting you: only the button's color changes.
• Adding existing projects decides by each folder's marker rather than by its name, and no longer offers the folder that contains them as a project.
• In Models you can remove a provider that was added from CleverCoder, a new agent's default model is the latest one, and the list syncs itself when the program opens.
• Each window has its own taskbar button, and its icon carries a dot with the color of the project's theme.
• You can reply to any card in the panel, not just questions, and the reply reaches the agent even if its terminal lives in another CleverCoder window. You can also write to an agent in another window, and the input bar tells you who you're writing to.
• Multi-question forms with a single submit: each question comes with its options and its own text box, and the ones you leave unanswered arrive marked as such.
• Cards draw tables, diagrams, and graphs in your theme's colors, show their links and open them in the browser, and animate as they enter, leave, and reorder, respecting the reduce-motion setting.
• Right-click on a card: copy, select all, and paste.
• Panel search finds matches in the whole history, from three letters, and tells you how many results there are.
• A button to view the conversation with a single agent, a bell to hide notifications, and a filter to see only what belongs to active contracts.
• Each agent tab shows with a colored bolt how much that agent is working, and whether it has the microphone open, the speaker on, or the antenna with its count of received messages.
• Every question and notice in the panel is born with its audio and its player: pause, seek, skip 10 seconds, and change the speed. Audio that didn't play is marked with a glow, and a question's audio reads out the numbered options so you can answer by ear.
• The voice screen is reorganized by what you want to do, and adds picking the dictation model, changing each agent's voice and pitch, previewing how it sounds, and Accept and Cancel buttons. Agents can install a missing voice, and if the configured one isn't installed the card says so instead of staying silent.
• Voice notes that CleverCoder prepares for WhatsApp come out in the format of WhatsApp voice notes and with their waveform.
• The roadmap shows all live work with a traffic light (green, yellow if there are unanswered questions, and red with its reason), each agent's queue with a lock to pin priorities, an icon per topic, and each contract's detail opening in parts. A contract can be put on standby with its reason, at the end of the queue.
• Automatic backup warns in the agent list when a destination fails twice in a row, and each terminal warns when the disk drops below 10 GB free and when it recovers.
• The Info screen shows the third-party components with their author and license, including the icon packs, and thanks the projects it builds on.
• Each message you write to an agent shows whether it arrived, is on its way or did not arrive, with a Retry button that never retries on its own.
• Agents can read other agents' panels, read-only, with filters by agent, type, state, contract, text and dates.
• The new-project wizard creates a leader with the name and model you choose (Opus by default), and its first session starts in plan mode.
• Create agent goes back to the list with the new agent highlighted, at the top of its group. The role catalog becomes a tab of the team panel.
• Info shows your registered email under the key and the days left in the trial, and lets you delete the key with a confirmation. After an update, a greeting confirms the key was detected.
• The team's global memory and each agent's personal memory are read and edited by section, with a lock and a version, so two agents don't overwrite each other.
• Each agent tab shows the tokens it used, and the project total adds them all up. The number is stored and only goes up: deleting a conversation file doesn't lower it.
• The Models editor shows three real states: available, retired with its date, or unchecked. Before, it marked every model as available, and 46 of NVIDIA's 103 were already retired. It is checked without spending tokens or using your key.

Changed

• Closing a terminal, or CleverCoder crashing, closes every process it launched: orphaned processes are gone.
• The background particles drop to 30 frames per second, to 15 when the processor goes above 70%, and freeze when the window loses focus.
• Card voice is generated one at a time across the whole machine and at low priority.
• The panels of tabs you can't see are suspended instead of using 64 MB each.
• Each open CleverCoder has its own web engine, so two windows no longer step on each other, and flat screens reopen in under half a second.
• The panel opens with fifty cards and brings the rest before you get there. Switching agents, filtering, replying, or changing theme no longer rebuilds it entirely.
• Computing usage reads 85% less disk per minute in each window.
• Messages between agents allow up to 7,000 characters, up from 4,000, and if they're trimmed the notice goes at the top.
• The program's Spanish texts, and the tools', now use neutral Spanish instead of a regional form.
• There is one panel and one typing bar per window: switching tabs switches owner without reloading, and it uses less memory. Each agent keeps its own draft, so what you type for one doesn't leak to another.
• Pressing an agent's tab only changes who you write to: the panel keeps showing all the chats, and filtering is done with the filters. The highlight moves on the same click, the terminal button takes you to the chosen agent's terminal, and its icon appears next to the name in the typing bar.
• With a filter on, the panel fetches only the messages that filter shows: before it fetched 400 and you saw between 70 and 117, now you see them all.
• The tab border and the little lightning bolt follow the agent's turn: with the turn open, at least green, and when it ends, back to zero at the same instant.
• Messages between agents wait for the agent to finish its turn and yield while you type. Auto-continue jumps on the same end of turn as the bell, and a change to its interval is applied live, without reopening the terminal.
• Entering a hidden terminal is faster: the retained output is processed at about 300 ms per MB, down from 525.
• The installer closes the leftovers of the previous program and updates without asking. If another program holds the old folder, it names it and changes nothing. Isolated agents and logs no longer live in the install folder.
• The repository and downloads move to github.com/clevercat64.
• The demo mode, the Terminal IA row of the agents panel and the button that turned a loose terminal into a leader are gone. The model and bypass are chosen in Create agent.
• Agents no longer ask you with terminal menus: every question reaches you through the panel.
• The agents' rules say where to save each thing (the state of the work, what was learned and what the whole team should know) and no longer contradict the guide they receive at startup: who builds in a team, keys go to the vault and test screens are drawn in the panel. That guide adds a diagram of each agent's home.
• 19 tool parameters explain what each value does.
• The project's Space tab is called Global memory, voice speed is shown as a multiplier (x1.123), and the saved-context card shows clean text with the time on top. The scroll bars of every page use the page's style.

Fixed

• Agents no longer get stuck when waking up: the system waits for them to finish starting before pasting their context.
• What you type no longer sends itself or gets split into pieces, and a long paste goes in right away instead of waiting 8 seconds.
• Clicking the "thinking" indicator no longer freezes the window, and that indicator turns off when the agent finishes, not on a timer.
• Git no longer hangs for up to 30 seconds when running inside the tools.
• The panel no longer erases what you're typing or moves you from your place when messages arrive, agents wake up, or you reply to a card.
• A message that arrives while a terminal is restarting is delivered on reconnect, and while an agent clears its context, messages from others wait and are delivered in order, without duplicates.
• An agent with an accent in its name no longer breaks its buttons, its message inbox, or its folder.
• With several CleverCoder windows open, agents no longer erase each other, and a new team with two or more agents no longer starts out broken.
• Claude Code's settings can no longer be emptied by themselves when the file is read badly.
• The terminal no longer starts out narrow and cut off when it opens from the team panel.
• Going in and out of Notes no longer leaves the window without its bars.
• Texts, fields, and borders are readable in all ten themes, and the agent panel is readable in the light themes.
• Auto-continue: a new project no longer copies another project's text, and a short text no longer stays typed without being sent.
• Closing a contract checks that it was saved, and a contract with a lot of conversation can now be read, because it's requested in parts.
• A failing audio is reported as a failure instead of showing as heard, no longer jams the queue, and two audios no longer play at the same time.
• A message between agents that isn't seen confirmed is no longer considered delivered without having gone in: it is looked for up to 15 seconds in the agent's conversation and, if it isn't there, it is pasted again once with an empty box and you are notified. The notice states the cause and asks you to resend only when the failure is final, without duplicating.
• A panel message to an agent with an open turn no longer ends up as "did not arrive", and Retry no longer duplicates it. Waking an agent no longer gives a false "did not go in" when the message sits in its queue, and a paste that didn't go in is no longer marked as injected.
• Party agents can see their own conversation history again: the wake-up check, the send confirmation and other detectors weren't working because they looked for it in the project folder.
• When an agent wakes up, the system checks that the message went in; if not, it retries once and, if it can't be confirmed, a notice appears in the panel.
• An agent's /clear waits for its turn to really end, and for the next one if another started from a queued message. Before, it went in mid-turn and the new session started without its context. A message from you or auto-continue no longer lands on top of the cleanup.
• No more double bell or double auto-continue jump when a turn ends.
• The panel no longer freezes the window when loading more cards, no longer pulses or shrinks when opening, and no longer gets stuck or restarts when scrolling up through history with a filter on. It doesn't stop showing old messages when you switch filters several times, and an already-answered question leaves the Questions filter.
• Switching agent tabs no longer waits: the panel read a 115 MB database three times per update, now it reads it once and all the panels share it (from 726 ms to 2 ms in tests). Each agent's pages start with the theme's background and the destination tab no longer hides.
• A card that asked for audio and stayed mute shows the generation error notice, and a voice command to an agent that isn't the visible tab's no longer leaves the window mute for 5 minutes.
• Going from Party, Pool or Contracts to Info or Settings no longer leaves the terminals covered, and a terminal whose web viewer crashed shows up again instead of staying in a 0-pixel row. Capturing a terminal works even if its tab is hidden, and closing a window's last tab frees its panel and voice.
• After a browser crash an extra typing bar no longer appears, the form answer boxes no longer have a double border on focus, and the floating agents column moves down with the context bar instead of covering it.
• The accent color and the name in notifications are readable in all ten themes: before, the accent failed in 5 of 10 and the name's contrast in a light theme was 2:1, now it is 4.65 or more. The delivery marks on messages are also readable in light themes.
• The thinking indicator no longer turns on when the program answers a permission by itself.
• Models loads the online catalog again; it never loaded because it repeated the same model across several providers. A favorite can follow its family and never moves to a retired model.
• Deleting the key no longer gets undone by the revalidation on open, and the memory tab and the instructions mirror no longer lose edits.
• Roles that are not in the database show their factory icon in Create agent and in the catalog, not the mask.
• The leader's messages to agents arrived as an inert block, and the tool that says who is working and since when returned an empty list.
• The contracts list with all states failed again. A short hour-and-minute number finds its task even if others start the same way, and if it is truly ambiguous it says so and lists the options. An invalid state names the parameter and lists the valid ones.
• Code search now finds struct, record, delegate, enum members and constructors, and sees properties, fields, events and methods used as a value. A member declared after a nested class no longer ends up as a child of that class.

Security

• The antenna listens only on the local address by default, and on the local network only if a terminal asks for it; the firewall rule is limited to that network.
• What comes in through the antenna is marked as unverified, and no agent can sign a message with another agent's name.
• OpenCode's NVIDIA key can be moved to the vault and is handed only to the OpenCode terminal. Environment variables are shown with their secrets hidden.
• "Set status line" is our own script that doesn't use the network or touch credentials: it only changes the status bar key in Claude Code's settings, with an atomic write.
• The DeepSeek and Kimi keys are also delivered from the vault, and only to the OpenCode terminal. After moving a key to the vault, the button offers to delete the backup that holds it in plain text, only if the vault stores the same one.
• The antenna's firewall rule drops the ports nobody listens on.

Added

• The conversation panel shows which question each reply answers, with a direct link to jump to it.
• Screen mockups can now be viewed directly inside the conversation panel, without opening a separate window, and any of their elements can be clicked to point at them.
• A button to jump to the end of the conversation panel appears only once you've scrolled far enough away, and fades out on its own.
• Panel tabs now respond to hovering the mouse over them, and an already-answered question is told apart from a pending one by the color of its border.
• AI engines you don't have installed stop being offered, both when you pick one yourself and when the AI spins up an agent on its own.
• Voice can save what it says to a file, in addition to speaking it out loud.
• The tools catalog now also sees the ones that live inside an installed plugin, with a notice explaining where they came from.

Changed

• The conversation panel updates instantly when a new message arrives, instead of reloading the whole thing every time.
• The panel no longer erases the text you're writing, and no longer jumps you to the end if a message arrives while you were reading further up.
• Team panel messages show who sent them and who they're addressed to, using the agent's real name.

Fixed

• The public changelog page (this very page) used to render without a title, subtitle, or footer.
• Several windows in the app were left with blank icons or stuck unclosable if the internal graphics engine crashed and restarted — they now recover on their own.
• With several terminals open, closing just one could trigger several duplicate notifications instead of one.
• The automatic backup size cap field now has a confirm button, shows updated usage instantly, and accepts up to 1024 GB.
• The list of folders the automatic backup skips now updates itself with each new version, instead of staying fixed from when you first set it up.
• The first terminal you open no longer briefly shows the prompt bar in the wrong color.
• Going back to the conversation from Notes no longer leaves the prompt bar hidden.
• Voice now warns when it reads text in a language that has no voice of its own configured, instead of silently reading it with another language's pronunciation.

Security

• Bug and feedback reports you send no longer include file paths from your computer or your Windows username.

Added

• Each agent can have its own auto-continue message, on top of the one the whole project uses.
• The terminal's card panel renders tables, bold text, and line breaks in their real format.
• Voice is listed in the tools catalog, with its name and description.
• Caveman Distillate comes preactivated for any agent, with nothing to install.
• Verdesk and Piumy come preactivated out of the box, ready to use with nothing to install. The title bar shows each agent's own icon.

Changed

• A contract with pending parts can't be closed by mistake, and the list groups each one with its subcontracts.
• Compiling clearly explains any error that keeps it from starting.
• The installer is lighter, and voice dictation is ready to use as soon as you install CleverCoder.
• The tools catalog shows the command to install each one, its full manual, its logo, and separates the built-in tools from the ones you added.
• Installing something from the catalog shows a progress bar and the reason if it fails.

Fixed

• Each project's card shows its real token usage.
• Kimi, Codex, and the other non-Claude engines open with their arguments and their own folder, just like Claude.
• On startup, a card shows the result of the boot checks.
• The card panel shows the name and icon of the agent asking, and the AI provider you chose.
• The card panel lets you type your answer freely, and you can see right away that it registered.
• clever-clear leaves the conversation's context clean.
• The agent's profile shows the model you chose.
• CleverCoder's own tools can't be turned off by accident.
• The project's file keeps its agents and settings even with two versions of CleverCoder open at once.
• Files that arrive through the antenna (cAPI) keep their real name.
• The antenna (cAPI) keeps the same setup even if you close and reopen the terminal.
• The project's automatic backup runs and verifies itself.

Added

• Licenses now carry a real expiration date, and the app respects it.
• A heads-up a week before your license expires, with a one-click renewal that doesn't lose work in progress.
• A clear notice the day the license expires, with a direct link to renew.
• Working offline no longer lets you keep using an expired license.

Changed

• Mouse and scroll now behave the same as any other program inside every terminal, with no more inconsistencies depending on which AI you have open.
• Copy and paste in the terminal no longer fight each other: right-click works directly.
• The startup banner of each terminal no longer flashes on open — it stays hidden while the program runs.

Fixed

• The info screen no longer shows "lifetime" to everyone: it now shows the real expiration date, or that it's expired with a link to renew.
• The purchase link points to the current store on all three buy buttons.
• Copying text no longer fails silently: it now warns if the clipboard was busy with another program.

Added

• Vault: an encrypted place inside CleverCoder to store your keys and passwords. Includes an approve-per-use mode for the most sensitive data, asking your authorization every time an agent needs to read it.
• The antenna now sends, not just receives. An agent can send messages and files to an external destination, and photos, audio or video that arrive through that channel are now saved to the project instead of discarded. Connection credentials survive closing the terminal.
• The Pool now shows the skills, plugins and MCP servers you can add to the project, and installs them with one click. It used to only show what was already installed.
• Icon picker expanded with four new libraries (Lucide, Tabler, Phosphor and Fluent UI), with a smooth grid even across thousands of icons, and icons that adapt to the active theme.
• Agent editor in the Party panel. Manage each agent's role, skills, working files and avatar directly from the panel, without opening its terminal.
• Support for Kimi and DeepSeek running inside Claude Code, and the Claude 5 family (Opus, Sonnet, Fable) added to the built-in model catalog.

Changed

• The app speaks consistent English throughout the interface, and now says "agent" instead of "mineral" everywhere it's shown.
• Redesigned the open-terminal window: a cleaner menu to resume a recent session, with unified colors between the open and resume actions.

Fixed

• The window no longer freezes when opening the contracts panel, and closing windows (icon picker, color wheel, file viewer, converter) now frees the memory they used instead of piling up over time.
• The mouse wheel scrolls the conversation up and down again, instead of replacing what you were typing with an earlier message.
• Copy and paste in the terminal, broken for a while, is fixed: text selects like on a web page, with a right-click menu for copy, paste and select all. Also fixed an emoji in an agent's name throwing off the surrounding text.
• The Build button no longer fails silently in some cases.
• Several stability fixes in Party mode (multi-agent).

Security

• Updated an internal component to close a known security vulnerability, with no visible change for the user.

Added

• Support for Ollama as an AI provider, and shortcuts for DeepSeek R1.

Changed

• The terminal bar groups the gear and antenna buttons into a single control, cleaner than before.
• The startup flags editor now has one tab per AI provider (Claude Code, Kimi, Codex, OpenCode, Gemini, Ollama), each with its own options.

Added

• Ctrl+Shift+C shortcut to copy the selected text in the terminal.
• Support for more AI providers in Party mode: Gemini CLI, Codex/GPT CLI and OpenCode CLI.

Fixed

• Several stability fixes in Party mode (multi-agent).

Added

• Automatic build with a direct link to the Run button. When you build, CleverCoder detects the project type, generates the artifact and updates the ▶ Run button on its own — no manual editing needed.
• Support for other AI providers in Party mode. You can now use Kimi, GLM, DeepSeek, OpenCode and other engines besides Claude for your agents.
• Antenna to connect external agents or devices to a terminal. Encrypted connection with a dedicated button, a live activity indicator, and configurable delivery delay and allowed attachment types.
• "Isolated" profile when creating an agent. Runs it in a separate environment, with no access to the real project code or system commands — meant for external or lower-trust agents.
• Work location picker when creating or editing an agent. Choose whether it works on the main project or on a new branch, with a name generator for that branch.
• Unified viewer for web pages and CUI designs. Includes a prototype-style navigation mode, one-click element selection, on-screen text editing and an Undo function.
• New color picker. Hue wheel, eyedropper to pick a color from anywhere on the screen, preset palettes and color-harmony schemes, in a window that doesn't block the rest of the app and respects the active theme.
• Edit files directly in the built-in explorer. Save changes with Ctrl+S while preserving the file's original format, and sensitive files (passwords, credentials, keys) are protected as read-only.
• Option to resume a previous conversation when opening a terminal. Instead of always starting from scratch, pick a previous session from a clear list.

Changed

• An agent's editor now lets you change its icon after creation too. The icon used to be set only at creation; now it can be updated anytime, just like the model or the role.
• The button to add a terminal to the multi-agent team now opens a clear dialog. Lets you choose between bringing in an existing agent or creating a new one (name, color and icon), instead of assigning an identity automatically without asking.

Fixed

• Automatic recovery when the terminal's built-in browser closes unexpectedly. No longer needs any manual action from the user to restore the terminal.
• Several stability fixes in Party mode (multi-agent). Messages that didn't arrive or took time to show up, identities and notifications getting mixed up between agents, stale icons and titles, and occasional failures starting or reopening agents.
• Fixed several panels and windows that showed mixed Spanish and English text.

Added

• File explorer with diffs, editing and saving. New window that shows which files changed (new, modified, deleted) with syntax highlighting, updates live as you edit, and lets you edit and save code right there without losing unsaved changes.
• Backup with multiple destinations and integrity verification. Recognizes your USB drives by identity even if Windows changes their drive letter, shows whether each destination is up to date or behind, verifies saved copies are intact, and lets you restore files from the panel.
• Checkable milestones in AI work contracts. Every contract now has a checklist you can tick off as it's completed, with a visible progress counter.
• Voice input and output in the conversation panel. Dictate your messages with the microphone (hold to talk) and have the AI read its replies out loud; the language adjusts automatically and the microphone mutes while the AI speaks to avoid echo.
• Wizard to create new AI agent roles. Create a new role without hand-editing files, with an icon picker built in.

Changed

• Clearer backup exclusion list. What the backup doesn't save (regenerable files like builds or caches) now shows in neutral gray instead of red, which looked like an error.

Fixed

• Cleaner, less intrusive pending-context notice. The warning icon no longer clips or shifts the other tabs, and the notification only appears on login or when clicked.
• More reliable sending of long messages. Text reinjected when reopening a project, or messages between agents on a team, would sometimes get stuck in the text box without sending; now it sends reliably even with long contexts.
• Closing the app no longer gets stuck. A hung shutdown can now be cancelled with a button or the Escape key.
• Sync across multiple open windows. A project created in one window no longer vanished from the list in the others, and switching the visual theme now applies everywhere.
• Switching your AI connection no longer closes every open session. Before, changing certain settings could suddenly close every open Claude Code window across every project; now it only offers to restart the affected session.
• Overall stability of the multi-agent team mode. Fixed several minor issues with icons, incorrectly shown model names, and syncing between agents on the same team.

Added

• New work Contracts panel. Every task you give the AI is now logged with its goal, scope and files involved, and automatically linked to commits; review, approve, close or revoke it from a dedicated panel, with an expanded window for reading long contracts.
• Active contracts can be renegotiated on the fly. If an improvement or a necessary scope change comes up mid-work, you can adjust the contract without closing it; the panel keeps a change history with before/after comparison.
• New conversation panel inside the terminal. Besides the classic terminal mode, switch to a chat-style view with messages, notifications and one-click answer options; it has its own zoom and keeps history between sessions.
• Per-project color palette. Each project can have its own color settings on top of the base theme, saved along with the project.

Changed

• The saved prompt library moved to the Tools menu, next to the MCP manager and the tool installer.
• Classic terminal mode lost its own prompt text box; you now type straight into the console, and that box lives in the new conversation panel.

Fixed

• Contracts panel unreadable in several themes. Text showed up black regardless of the active theme; it now follows the chosen theme's colors.
• Recovery when the terminal failed internally. If the terminal stopped responding (for example, after waking the computer from sleep), the AI session became unreachable; now a notice appears to recover it without losing work in progress.
• Bug reports and feedback now actually arrive. A configuration issue meant no report sent from the app reached the development team.
• More reliable recovery of your last session when reopening a project. Fixed several issues where the "continue where you left off" notice wouldn't appear, would repeat unnecessarily, or would paste corrupted text.
• Order of project cards on the home screen. Sorted by most recent first again.

Added

• Simplified new-project wizard. Now you just enter a name, a folder and press start; the rest of the planning happens in conversation with the AI the first time you open the project.
• Version-control options when creating a project. Choose a local repository, no version control, private GitHub, public GitHub, or decide later.
• Per-project skills catalog. Travels with the project and survives backups or migrations; mark skills as required, disable the ones you don't use without losing them, and update the catalog from the global library whenever you want.
• Tabbed CLAUDE.md editor. See and edit the project's file, the one from parent folders, and your global user file, all in one window, with automatic refresh if they change on disk while you work.
• New open-terminal window. With cards for plain Claude, the lead agent and each agent on your team, plus a random name generator for creating new agents.
• Agent editor. Change the model (Opus, Sonnet or Haiku) and the role assigned to an agent after creating it, with no need to recreate it.
• Close dialog with open terminals. When closing the app, choose to close agents cleanly or force-quit, instead of having them cut off without warning.

Changed

• Automatic agent team setup. The multi-agent team now assembles itself quietly the first time you open the lead or an agent, with no dialogs or manual steps beforehand.
• Direct model selection when creating the project. Choose Opus or Sonnet from the wizard, with an automatic suggestion based on your subscription plan.

Fixed

• Loss of the project list. Fixed an issue that could wipe it after an unexpected disk-space problem; it now saves more safely and recovers itself from a backup copy if something fails.
• Skills wiped when updating the catalog. Updating the skills catalog no longer deletes the ones you'd added manually.
• Contrast and readability in dark and colored themes. Fixed several dropdown menus and buttons that showed up with poor contrast.
• "Plain Claude" terminal inherited someone else's context. Fixed it sometimes still starting with an agent's context or identity instead of a clean start.
• Stability when creating, closing or reopening team agents. Several fixes reduce duplicates and improve recovery when something gets interrupted midway.

Removed

• Rarely used editors with an unfinished interface. Removed the visual team editor and the role picker on opening a terminal; those actions are now done with chat commands.

Added

• Pixel-art office view for Party mode. Each active agent shows up at its own desk, which glows while the agent is thinking, shows messages traveling between desks, and turns red when there's an active alarm.
• Agent cards with live status. Show who's active or asleep, let you wake an agent with one click, and remove it entirely if it's no longer needed.
• Recent activity panel next to the office view. Gathers decisions, builds, messages and other events from each agent, with filters and direct links to related documents.
• Floating prompt bar, independent of the main window: draggable, resizable, and remembers its position between sessions (even across multiple monitors).
• Per-agent alarm system. Schedule an agent to get a message at a fixed time, on a recurring schedule, or when a file changes or a new message arrives.
• Keepalive mode. Keeps a tab active during long waits so it doesn't disconnect.
• Extra panels in the office view: project roadmap, active tasks and feature status, all visible at a glance.
• Checkbox to opt out of usage telemetry, available in Settings.

Changed

• Backups are now generated as a single compressed file, with an incremental option (only what changed since the last one) and a progress bar showing which phase it's in.
• Auto-continue now works independently per agent. Turning it on for one agent no longer starts the automatic cycle on the others.
• You can now open several CleverCoder windows at once, for example to work on two different projects side by side.

Fixed

• Fixed several Party mode stability issues: messages that didn't arrive between agents, agents showing as active when they weren't, tabs freezing or appearing blank when invoking a new agent, and errors opening the office view.
• Better performance with many tabs open at once; the interface no longer slows down.

Added

• Costumes: specialized AI agents that work as a team. Open helper agents with defined roles (programmer, reviewer, supervisor, among others), each in its own tab, coordinated within the same team.
• Visual costume picker. A dedicated window to choose the helper's role, the AI model, and write an initial task before opening it.
• Redesigned team button. Invite terminals or projects, dispatch tasks and check the team agreement's status at a glance.
• "Send task" shortcut on every team member. Send a message or task to a teammate directly from its card, with no need to type commands.
• Visual AI activity indicator. An on-screen particle effect reflects in real time how hard the assistant is working.

Changed

• The team message counter separates direct messages from broadcasts, so you can tell at a glance how many of each there are.
• The Notes and Send buttons at the bottom of the terminal adjust their icon and label depending on whether the notes panel is open, taking up less space without losing clarity.

Fixed

• Fixed a freeze when opening a costume (AI helper) that could unexpectedly close the window.
• Restored the window's title bar to its correct look: rounded corners, shadow, and terminal dragging working normally again.
• Fixed several stability issues in agent coordination: duplicate tabs when opening a helper, tasks landing on the wrong tab, and incorrect identity data with several tabs of the same project open.
• Fixed regional-slang text to keep neutral, consistent Spanish throughout the app.

Added

• New team feature: Party. Connect several projects into one working team and see it represented as a pixel-art office, with a character per member and cards showing their status, pending messages and token usage.
• Internal messaging between team members. Send direct messages or broadcasts to the whole team from a built-in inbox, with a sound notification when something new arrives.
• Configurable team participation rules. Whoever organizes the team can password-protect it, set a token limit per member, and define what information is shared, all from a dedicated panel.
• Initial setup wizard. The first time you use each new version, CleverCoder detects if a needed tool is missing (like Git, Node.js or the browser component) and walks you through installing it, checking progress in real time.
• Available-update notices in the Tool Installer. Tools with a newer published version now show a badge and a one-click update button.
• Notes panel reorganized into tabs. Besides your notes, browse the project's .md files, your archived notes and all documentation from the same panel, with drag-and-drop reordering and a high-contrast reading mode.

Changed

• The built-in terminal prefers PowerShell 7 when it's installed. This speeds up chained commands and avoids the old-version update notice.
• Auto-reply, auto-continue and notification-sound preferences are now saved per project, instead of sharing one global setting.
• Background terminal tabs use less memory. They suspend automatically when out of view, without losing history or the active connection.

Fixed

• The floating terminal bar could disappear after the computer woke from hibernation. It now restores itself automatically.
• Fixed several intermittent freezes opening certain new panels, including occasional failures starting the app with uncommon prior configurations.

Removed

• Removed the Win95 visual theme from the theme picker, and Visual Studio Code is no longer available in the Tool Installer's one-click catalog.

Fixed

• Small internal startup fix.

Changed

• Visual and performance improvements across the general interface.

Added

• Built-in automatic update channel in the app.
• Improvements to the icon menu and the floating sidebar's behavior.

Changed

• Style adjustments so buttons and panels follow the active theme.

Added

• New buttons on the terminal bar to change font size and hide the bar in tab mode.
• Auto-hide for the floating bar after a few seconds of inactivity.
• "Recommended / All" icon filter based on project type.

Changed

• Full system icon catalog expanded to over a thousand glyphs.
• Better visual integration with each color theme.

Fixed

• Various fixes in the icon-pack preview.

A long cycle with lots of improvements to the Icon Picker, Notes and the terminal bar. Summary:

Added

• Icon Picker with several systems at once: system fonts, popular icons and downloadable color emoji packs.
• Automatic emoji-pack download on first use, with a non-blocking progress bar.
• Bilingual (ES/EN) name search plus keyword search, tolerant of accents.
• Variant popup (light/dark, filled, skin tones) when picking an icon.
• Emoji-pack management panel in Settings (download, remove).
• Per-project "receipt"-style summary: sessions, time and tokens, computed locally.
• Notification sounds section with 24 free effects (Kenney), theme-based selection, custom sound or system sound.
• Visual activity indicator at the bottom of the terminal (color pulse tied to the AI's thinking rhythm).
• Precise detection of when the assistant finishes replying: the bell only rings at the real end, not on mid-reply pauses.
• Dual Notes/Send button at the bottom of the terminal with a pending-notes counter.
• Drag to reorder notes; renaming, deleting and creating notes reflects instantly.
• Notes panel with text styles, color, tables and shortcuts.
• ▶ Run button with a visual heads-up when there are changes ready to test.
• Two-step new-project wizard: pick a template and skills, then write the initial prompt.

Changed

• Reorganized terminal bar: the most-used buttons sit closer together, with consistent spacing.
• Every icon pack shows its license and attribution under its name.
• Notes now save as independent files named after their title (no hidden index).
• CleverCoder bundle in the wizard: one click installs the default skill set.
• Third-party skills visually separated from official ones, with links to their repositories.

Fixed

• Several focus issues between the terminal and floating windows (icons, ruler, rectangle, etc.).
• Emojis pasted into Notes with unwanted prefixes.
• Popups that stayed open or landed in the wrong position when moving the app.
• The Send button's counter is now correct from the first second.
• Colors and fonts that didn't update when switching themes live.

Added

• Icon Picker with a palette taken from the active theme, GPU acceleration, and shortcuts to paste the icon straight into the terminal.
• Local bundling of the main icon font sources (100% offline).

Changed

• Compact prefix:name icon format (easier to use in prompts and layouts).
• icon-picker skill so the assistant respects the chosen icon system.

Fixed

• Rendering fixes for several glyphs and a safe close for the picker.

Fixed

• The Icon Picker now pastes the code into the terminal correctly, end to end.

Added

• CUI 1.1 protocol — you can declare which font renders each icon, so the visual editor and the app look identical.

Fixed

• Adjustments to the build button to show a clean, consistent hammer icon.

Added

• Public skills catalog on GitHub: the community can propose skills via Pull Request.
• A single skills file serves both the wizard and the admin panel.
• CLAUDE.md editor using the same flow as the new-project wizard.
• Multi-language infrastructure with external JSON files (add translations without recompiling).

Changed

• Two visual skill sections: official CleverCoder skills (green border) and third-party.
• Duplicate skills from the same repository collapsed into one.

Fixed

• Duplication when checking/unchecking skills in the wizard.

Added

• New-project wizard built into the main panel.
• Automatic paste of the initial prompt after creating the project.
• Backup with local or remote-server support (username and password, or key), with no limit on destinations.
• Progress bar and automatic retry for locked files during backup.

Changed

• Simplified backup system: now simpler and more reliable, a straight mirror of the folder.

Added

• Layout editor with batch auto-send and a visible notice when outdated changes get discarded.
• On reopening, the editor recovers what you were building.

Security

• Image URLs are validated before use in the layout editor, for extra safety.

Removed

• Cleaned up unused parts of the editor.

Added

• Visual layout editor with real-time preview.
• CUI protocol to describe interfaces compactly.
• Import an image so the assistant can generate the initial layout.
• Measuring tools (ruler, rectangle) and direct color capture from the desktop.
• Snip: crop a screen area and paste it as a reference.

Changed

• Unified workflow: the assistant replies with a file that updates the preview automatically.

Added

• 20 new skills across 6 categories: Productivity, Development, Security, Organization, Content and Notifications.
• Direct install of skills and dependencies from the interface.

Added

• Mode, Effort and Model buttons on the terminal bar, with a quick switch to the mode you want.
• Quota bar widget (daily and weekly) with a tooltip.

Changed

• Simplified colors: green, red and gray by status.

Added

• 8 cyclable visual themes (Dark, Light, Forest, PinkLotus, DarkFlame, Ice, Matrix, DarkSakura).
• Color Picker in the terminal: pick a color from the desktop and it pastes as hex.
• Title bar colored to match the active theme.

Changed

• The app is now called CleverCoder.

Added

• 6 cyclable themes, a project counter in the title bar, quota architecture split into a quick check plus a detailed background check.

Fixed

• Exact calculation of quota-reset cycles.

Added

• /compact button with a live context-usage bar.
• Continue button that reloads the project's context from its key files.
• Voice dictation shortcut (Windows Voice Typing).
• Account manager (Claude, GitHub, HuggingFace).
• Expanded prompt library with emojis and preview.

Changed

• Simplified terminal bar: the manual command field is gone.

• Project launcher with search, add, scan, rename and delete.
• Visual CLAUDE.md editor with 6 templates.
• Prompt library: save, tag, search, import and export.
• MCP server manager.
• Tool installer (winget).
• Per-project notes.
• Settings: theme, language, terminal and Claude path.
• Dark / Light / System themes, switchable live.
• English and Spanish, switchable at runtime.
• Built-in terminal.
• Auto-yes for Claude Code confirmations.
• Maintenance popup (summary, commit, backup, changelog, pending items).