Every desktop release, newest first. Apple Silicon macOS, Windows, and Linux.
v7.3.2Current2026-10-25
In Claude Code, each agent now has its own Computer Use REPL. The main agent and its subagents, such as the CodexUse operator, no longer wait for each other's calls, and one agent's timeout no longer wipes another's variables. In 7.3.1 that waiting took 15% of the time agents spent in Computer Use, up to five minutes for one call. Computer Use updates the Claude Code entry on its own when the app starts.
Agent tabs close again once the agent is done. After you showed the agent browser once, for example from its Dock icon, no agent tab closed for hours; now only an agent that shows the browser for a sign-in keeps the tabs of that request. Tabs of an agent that quit before it could clean up now close within a second, and after the Mac sleeps an agent keeps its tabs instead of losing track of them while they stay open.
A long wait no longer runs a call past its time limit, which reset the agent's work: waits stop just before the limit, and a call that sleeps longer than its limit gets a longer one.
Clearer errors for agents: reading navigator, localStorage or cookies inside a page read, opening chrome://, file:// or data: pages, writing two values at once, and searching for an element by the wrong role now each say what to do instead. The operator also knows how to read files and save screenshots, and an agent doing website work itself is reminded every 25 calls to hand it to the operator.
Computer Use works with ChatGPT 26.1021.31345.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
A page that takes a while to finish what a click started, such as generating an invoice or a report, now gets to finish it. CodexUse took such a page for a stuck one and stopped its script, so the click did nothing and the agent tried again and again. While the page is still busy, the agent is now told to wait and read it again instead of being told the element is missing.
When an agent's selector matches nothing on a page, the error now lists the elements most like it, as ready locators, so the next call usually works. Selectors that matched nothing were the most common agent error in our week of real use.
In Claude Code, an agent or subagent that keeps doing website or app work itself is now reminded, at its third and tenth call, to hand the rest to the CodexUse operator, which works on Sonnet in a short context. Computer Use updates the Claude Code entry on its own when the app starts.
Repeated work becomes a recipe more often: an agent that runs the same long flow several times in one request is asked to save it, and a recipe that does nothing no longer stops the reminder.
Clearer errors for agents: a call that ran past its time limit says what was lost and how to avoid it, blocked raw CDP commands name the call to use instead, and a search of an app's state that found nothing points to axFind.
With analytics on, the daily summary now includes Computer Use health per CodexUse version, such as error and slow-call rates in ranges, so we can tell whether a release made Computer Use better. It sends no pages, code, URLs or recipe names.
The macOS app now runs on Bun 1.4.4. Computer Use works with ChatGPT 26.1016.61520.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.3.02026-10-21
In Claude Code, website and Mac app tasks of more than two steps now go to the CodexUse operator, a subagent that runs on Sonnet in its own short context. Each browser or app step used to run on the main model with hundreds of thousands of tokens of context; the operator's steps use 7,000 to 25,000. To install it, click Update next to Claude Code in Computer Use, under Add to your agents.
Repeated tasks now take one call. Agents save a flow that works as a recipe in ~/.codexuse/recipes, and the next request runs it in one call; in our tests a release wizard that took more than eight calls ran in one call of 34 seconds. When tasks on a site keep taking eight or more calls, the agent saves a recipe on its own, without you asking. A recipe step that sends, posts or deletes runs only when your request asks for exactly that, and a recipe can wait up to an hour for a slow page, such as a long answer, without the agent checking again and again.
A page whose script stopped answering no longer makes every call on its tab wait 30 seconds, and its tab closes again. After the Mac woke from sleep, CodexUse started more copies of the agent browser's helper processes each time; it now keeps the ones that run. When the hidden agent browser stops answering, CodexUse saves a sample of Chrome's threads beside its log, so the cause of a hang can be found.
For agents: waitFor() counts an element that shows as found on heavy pages such as Gmail settings, tabs from browser.tabs.get(), new() and selected() get the same helpers as other tabs, and an action given a timeout over 3 seconds waits for its element first.
A mistyped license key now says that the key could not be verified, instead of asking you to try again later.
Computer Use works with ChatGPT 26.1016.51333, which CodexUse copies the next time an agent starts.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.2.42026-10-20
Providers now lists OpenCode Go, so you can add the GLM, Kimi, DeepSeek, and MiMo models of your OpenCode Go plan without typing its address. Requests to the address in OpenCode's docs now also send the session ID that OpenCode Go requires.
When Account Pool refreshes the sign-in of the account Codex is signed in to, Codex now gets the new sign-in too. Before, Codex could later ask you to sign in again.
Account Pool now bundles CLIProxyAPI 8.0.35. It refreshes Codex sign-ins 4 days before they expire instead of 1, refreshes right after a rejected token, and no longer lets an older copy of a sign-in replace a newer one.
The macOS app is now built with Electrobun 2.0.4, the stable release of the 2.0.3 beta, and still runs on Bun 1.4.3.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.2.32026-10-17
When an agent shows the agent browser and asks you to sign in, its tabs now stay open while you finish. In Claude Code the agent's question ended its reply, and that closed its tabs, so the agent came back to tabs that were gone. Tabs stay while the browser is shown, also when you show it yourself.
For agents: the first tab of a session no longer prints the page's state twice, and element ids passed as text, as in tab.click("7"), now also work in browser tabs. A click, fill or check that timed out is no longer tried again on its own, since its input may already have reached the page. A failed call says how many clicks, key events and navigations reached the agent browser, so the agent reads the page before it repeats an action.
Computer Use call logs on your Mac now also count what each call sent to the agent browser, without URLs or page content.
Account Pool now bundles CLIProxyAPI 8.0.32. It reads Codex streams more reliably, accepts Codex tool schemas with out-of-range number limits, retries closed HTTP/2 connections, and keeps a model's cooldown when a late request succeeds.
The macOS app now runs on Bun 1.4.3, built with Electrobun 2.0.3 beta.
Computer Use works with ChatGPT 26.1013.40934, which CodexUse copies the next time an agent starts.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.2.22026-10-16
The agent browser now quits when no agent uses it and no longer holds up a restart or shutdown. In earlier builds the helper that quits an idle agent browser failed to start, so its Chrome kept running hidden, and at a restart it agreed to quit but stayed open until it was force quit. CodexUse now also quits it itself when the Mac logs out, restarts or shuts down.
The agent browser stays behind your apps. Agent tabs open in one window that stays between replies, without taking focus, and only your own click on its Dock icon or window brings it forward; key presses, Cmd-Tab and clicks meant for other apps no longer do. When Chrome still brings itself forward, as when a page opens a popup, it is sent back within about 150 ms.
Show agent browser and Hide agent browser are now also in the tray menu and in codexuse-cua --show-browser and --hide-browser. A shown agent browser opens its windows on screen instead of in the corner and stays until you hide it, and Hide keeps the tabs you opened.
For agents: the Computer Use documentation now goes to files that the first result names. Claude Code moved results over 50,000 characters to a file and showed only their start, so agents missed the page state and errors of their first call. Results now stay under 45,000 characters and errors under 10,000, with the error first. Raw DevTools calls read their required documentation on their own, a closed tab's error names its last page, and more errors explain the next step.
Computer Use works with ChatGPT 26.1013.31256, which CodexUse copies the next time an agent starts.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.2.12026-10-14
Computer Use works with ChatGPT 26.1007.71326 and its new Computer Use runtime, which CodexUse copies from ChatGPT the next time an agent starts. Confirmation policies up to 20,000 bytes, the new runtime's limit, now reach agents. When the WebMCP settings in ~/.codex/browser/config.toml block a site's tools, the agent now learns where that setting lives, since CodexUse has no Site tools screen.
Pages that save files through the browser's Save dialog, such as a whiteboard export, now save through a normal download in the agent browser. Before, the dialog opened in the hidden window and the agent's call waited until it timed out. The agent browser restarts once to apply this, when no agent is using it.
For agents: element ids passed as text, as in tab.click("12"), now work. Failed calls also explain an action that got no element id or point, usually from a lookup that found nothing, and an app that could not complete an accessibility request.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.2.02026-10-13
Agent browser tabs now close after 15 minutes without use, even while background subagents keep a Claude Code reply open. Tabs an agent marked to keep stay open, and an agent that returns to a closed tab is told to open the page again. A reset during a reply no longer leaves its tabs open.
Computer Use now installs the OpenAI release that ChatGPT's own updater receives, so builds still in staged rollout are no longer installed early. Agents also get clearer hints for the errors they hit most, learn that typeText presses Return on each newline, and keep each call under 100 seconds so Claude Code does not move it to the background.
Model Providers is now Providers. OpenRouter shows every model it publishes instead of stopping at 256, the DeepSeek model test passes, and reasoning effort follows DeepSeek's and OpenRouter's current documentation, including max on OpenRouter. Current Codex versions no longer show a new-protocol warning on every model, and Together AI uses its new address.
An account named OpenAI, Azure, or Amazon Bedrock no longer makes Codex treat it as its built-in provider, which broke long conversations.
Account Pool now bundles CLIProxyAPI 8.0.27. It saves refreshed sign-ins safely and keeps a rate-limited account's other models listed.
The Codex config CodexUse creates and the YOLO preset now use gpt-6.1-sol. If you applied YOLO before, apply it again to switch.
Text across the app is shorter and says plainly what each feature does. Computer Use gives Claude Code, OpenCode, Cursor, and other coding agents the same Computer Use as ChatGPT and Codex.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.1.12026-10-12
The agent browser now opens every agent tab in one Chrome window instead of one window per tab. Tabs behind the front tab keep working, so several agents share the window, and an agent no longer switches away from the tab you are looking at.
In Claude Code, agent browser tabs now stay open for the whole reply. Before, two minutes spent on other tools closed them, and the main agent finishing its reply closed the tabs a background subagent was still using.
A page script that keeps running after its command timed out is now stopped, so the next screenshot or click no longer waits behind it. Raw DevTools access no longer stays blocked after a page load that Chrome dropped.
For agents: hover() scrolls the element into view first, scrollIntoViewIfNeeded() works, a wait the code never awaits no longer resets the REPL and its variables, and getApp() with a bundle id that several copies of an app share opens the running copy.
Failed calls explain more causes and the next step: closed tabs and how to keep them, what playwright.evaluate cannot do, the 3-second default of raw CDP commands, file choosers that did not open, and variables lost after a reset.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.1.02026-10-12
Every account now follows your own ~/.codex/config.toml. Change the model, approvals, sandbox, or MCP servers there, and each account picks it up the next time it starts. A value an account changes for itself, such as a theme picked inside its Codex window, stays on that account. Logins, API keys, and custom-model routing are never copied between accounts.
Settings > Codex defaults replaces the TOML editor. It shows the file with Open and Show in Finder, a summary of the model, reasoning effort, approval, and sandbox, and each account that differs, for example Differs: desktop.theme, with a Reset button. If the file has a syntax error, it names the line and every account keeps its current settings.
History and continuing a task no longer stop with "changed outside CodexUse" after Codex writes to an account's settings. On the first launch, each account's current settings become its own values, so nothing changes until you edit ~/.codex/config.toml. Shared defaults that differed from that file are kept as defaults-before-v13.toml.
The codexuse command now ships inside the app. In Settings, Add codexuse to Terminal puts it on PATH for zsh, bash, fish, or the Windows user Path. It runs on the app's own runtime, needs no Node.js, and always matches the app's version. The in-app npm installer is gone. The npm package codexuse-cli remains for machines without the app and now needs Node.js 22.13 or later.
Removed from the CLI because the app does them: codexuse best, codexuse profile autoroll, the Codex window flags of profile switch, and --runtime and --state-dir on account-pool. Each now prints what to use instead. codexuse profile switch now changes only the account a plain codex uses in a terminal, and refuses while the Codex app has ~/.codex open.
The CLI and the app now read and write Account Pool keys and sessions through the same code, and an older version refuses to rewrite Account Pool data saved by a newer one instead of dropping its sessions.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.0.12026-10-11
An OpenAI account opened with Open in Accounts now gets the CodexUse status pill in its Codex window right away. Before, only windows opened from the pill or from a custom model account had it.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v7.0.02026-10-11
CodexUse 7.0 focuses on one job: all your Codex accounts, side by side on your Mac, each in its own window of the official Codex app.
ChatGPT Web accounts are removed. On the first launch, CodexUse stops the ChatGPT Web helper, deletes its data, its account folders, and its saved keys, and shows a one-time notice in Accounts. Your OpenAI accounts, including one you linked during ChatGPT Web setup, stay. Three things live outside CodexUse, so remove them yourself: the CodexUse Chrome extension, the OpenAI tunnel API key you created, and the ChatGPT connector you added.
Cloud Sync is removed, and the encrypted copies stored on the CodexUse server have been deleted. Accounts on each computer are unchanged. The codexuse sync command is gone, and older versions that try to sync now get a message that Cloud Sync was removed. CodexUse also deletes the remembered sync passphrase from this computer.
The sidebar is simpler: Accounts at the top, a For agents group with Computer Use, Model Providers, and Account Pool, and Settings at the bottom. Settings is now one page with Preferences, Codex defaults, the CLI, and Support. Links to the old Settings sections open Settings.
The in-Codex overlay keeps the status pill with the account switcher, Auto-roll, and Stay or Switch now. Wide conversations, Remember position, and Jump to user turns are removed, and a Codex window that is already open drops them the next time CodexUse updates it. Refill timers now start only automatically; the Start now button is gone.
Anonymous usage reports now say where an account switch started (app, menu bar, Cmd+K, the Codex overlay, or Auto-roll) and include a once-a-day summary of rounded counts for Account Pool requests, custom model requests, and Computer Use calls. They never include account names, models, URLs, or window titles.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Codex windows and Computer Use for agents need macOS; on Windows and Linux, CodexUse keeps accounts and custom models separate for the Codex CLI and Account Pool. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v6.7.02026-10-11
The desktop app has a new design. Accounts, Settings, and the other screens now sit in a sidebar window with updated controls, icons, and light and dark themes. The website uses the same design.
Computer Use for agents now opens websites in the CodexUse agent browser: Google Chrome with its own profile, running in the background with its windows off screen, so it does not take focus or touch your own tabs. Import cookies from Chrome, Helium, Brave, Edge, Arc, Vivaldi, or Chromium in Settings → Computer Use so agents start signed in. Agents no longer use the ChatGPT browser extension.
Each agent session can show and hide the agent browser, and the browser quits a minute after its last tab closes. The agent browser supports page WebMCP tools, handles dialogs, tabs, and downloads more reliably, and refuses downloads the runtime did not request. When ChatGPT Desktop has a matching Computer Use runtime, CodexUse reuses it.
Account Pool now uses the unmodified CLIProxyAPI native engine instead of app-server turns for Responses and Chat Completions. Builds and each new engine process resolve the latest stable release with official checksum verification; verified bundled or cached binaries remain usable offline with a warning, and running requests are not restarted for updates. Streaming, function-tool round trips, raw response continuation, and stateless store:false requests use the API path. Stored streamed responses are resumable at response.completed without waiting for EOF. Existing keys and routing settings are preserved; old app-server sessions are reset.
Updated workspace dependencies, including the API server library. This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows on ARM uses the x64 package through emulation. Computer Use for agents needs macOS. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v6.6.12026-10-05
The desktop now uses a compact toolbar, grouped navigation, and neutral light and dark themes. Accounts, Settings, account setup, task search, dialogs, and license screens share updated layouts and controls.
On macOS, CodexUse can recognize an existing managed Codex window again after the Codex background service restarts, including windows whose debug port was assigned automatically. Recovery verifies the original process and account folder before reconnecting.
Support diagnostics now include Codex window and connection status to help investigate account-window failures. Account names and folder paths are excluded from these fields, and error text uses the existing credential redaction.
Updated workspace dependencies, including the API server library. This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows on ARM uses the x64 package through emulation. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v6.6.02026-10-02
The desktop interface has been redesigned across Accounts, all nine Settings sections, onboarding, license screens, dialogs, and the Chrome Companion. Light and dark themes share clearer text, larger controls, consistent spacing, and layouts that work in narrower windows.
Accounts groups OpenAI, ChatGPT Web, and custom-model accounts with clearer usage and connection details. Auto-roll controls now sit below the account list. Account and provider setup shows progress, searchable model lists, and clearer loading and error states.
Find local tasks from the search button or Cmd+K. The separate History tab is removed. On macOS, stopped tasks can continue on another eligible account and model while preserving the original task; the destination is checked before a continuation is created.
Computer Use for agents on Apple Silicon macOS now downloads and maintains its own verified OpenAI runtime, without requiring ChatGPT Desktop or Codex Desktop. Browser setup reports extension and connection state separately, and Hermes Agent joins the supported agent setup options.
Follow system keeps tracking macOS appearance after leaving Preferences. Switching accounts, settings sections, and setup steps starts the new view at the top. Cloud Sync uses Upload and Restore labels, and configuration editors show unsaved changes and prevent duplicate saves.
This release includes macOS Apple Silicon, Windows x64, Linux x64, and Linux arm64 packages. Windows on ARM uses the x64 package through emulation. Windows updates remain manual, and unsigned Windows packages may show a SmartScreen warning.
v6.5.02026-09-28
Work is removed. CodexUse no longer runs planner and worker accounts, delegated runs, or the review and accept step. History, which used to sit inside Work, is now its own tab next to Accounts and Settings, and still finds every local Codex task from every account.
Each account's Codex settings moved from the Work panel to a Codex settings section on the account. It keeps the same choice between the shared defaults and the account's own override.
Existing Work worktrees stay in the CodexUse app data folder under work/worktrees, and codexuse/job branches stay in your repositories, in case one holds a result you never accepted. Delete them when you no longer need them. CodexUse removes the planner's delegation tool from that account's Codex config the next time it writes the config, for example when it opens the account's window.
Settings → Computer Use lists the coding agents found on this Mac and adds the server to the ones you pick. CodexUse adds Claude Code, Amp, and OMP itself, and uses add-mcp's list for others such as Cursor, OpenCode, VS Code, Windsurf, Zed, Gemini CLI, Goose, Cline, GitHub Copilot CLI, Kiro, and Claude Desktop. Agent entries now run the launcher ~/.codexuse/bin/codexuse-cua, so they keep working after CodexUse moves or updates.
This release ships for macOS, Windows, and Linux. Computer Use for agents needs macOS. The Windows package is unsigned, so Windows may show a SmartScreen warning, and Windows users install it manually because automatic updates stay off on Windows.
v6.4.22026-09-28
Computer Use for agents no longer asks you to approve each app or site. CodexUse approves the apps and sites an agent asks to use and tells the Computer Use runtime in ChatGPT Desktop to keep those approvals. Before, an approval lasted only until the agent paused for two minutes, so Claude Code asked about the same app again and again. A site's sign-in challenge still comes to you, because only you can sign in.
Agents act with your macOS permissions and no longer stop for an approval, so add the Computer Use server only to agents you trust. Settings → Computer Use says this below the agent list.
This release ships for macOS, Windows, and Linux. Computer Use for agents needs macOS; Windows and Linux get the same version with no other changes. The Windows package is unsigned, so Windows may show a SmartScreen warning, and Windows users install it manually because automatic updates stay off on Windows.
v6.4.12026-09-27
Windows: CodexUse no longer gets stuck restarting its background service on startup. Saving settings asked Windows to flush a folder, which Windows refuses, so installs with accounts failed the same way on every start. If the background service still cannot start, CodexUse stops after a few attempts and shows the error instead of retrying forever.
A Work setup failure no longer stops the rest of the app. Work screens show the reason, and switching or opening an account retries the Work setup on its own.
CodexUse reads the signed-in account from ChatGPT Desktop 26.924 and later again, which moved the connection to its Codex app-server.
Diagnostics in error reports no longer include your Windows or Linux user name in file paths. They keep the file name and line number needed to fix the error, and the names of failed startup steps are no longer hidden.
This release ships for macOS, Windows, and Linux. Windows and Linux users also get the 6.4.0 changes that apply to their platform. The Windows package is unsigned, so Windows may show a SmartScreen warning, and Windows users install it manually because automatic updates stay off on Windows.
v6.4.02026-09-27
Computer Use for agents lets Claude Code, OMP, and other MCP agents use Codex Computer Use from ChatGPT Desktop on macOS. Agents control Mac apps, and Chrome, Edge, or Brave through the ChatGPT extension, while their own model decides every action. CodexUse asks before an agent controls a new app or site.
Settings → Computer Use checks the setup when it opens, turns on browser use, and adds the server to Claude Code or OMP in one click. The installed app keeps those agent entries pointed at itself after it moves, and copies a config for any other MCP agent.
CodexUse recognizes ChatGPT Desktop 26.924 and later again, which moved its bundled Codex CLI.
This release ships for macOS only. Windows and Linux downloads stay on 6.3.0, which has the same features on those platforms.
v6.3.02026-09-20
CodexUse desktop now ships for Windows 11 x64 and Ubuntu 24.04 on x64 and arm64, alongside Apple Silicon macOS. Windows on ARM installs the x64 package and runs it under Windows emulation. Each package is built and smoke-tested on its own operating system and architecture. The Windows package is currently unsigned, so Windows may show a SmartScreen warning during installation.
Accounts, ChatGPT Web accounts, custom-model accounts, Account Pool, headroom and refill timers, History, tray, notifications, file and browser actions, and encrypted Cloud Sync work on all three platforms. Each account keeps its own isolated Codex home, and user data follows the platform location: Application Support on macOS, %APPDATA% on Windows, and $XDG_CONFIG_HOME on Linux.
One Codex window per account, planner-first Work runs, Auto-roll, and Telegram remote drive OpenAI's official Codex app, which ships for macOS, so they remain macOS features. Outside macOS those controls are disabled and say why; existing worker runs stay available for review and recovery.
Automatic updates are disabled on Windows because the packaging runtime cannot apply one reliably yet, so a newer Windows package is installed manually. Linux update artifacts ship with the release and the full upgrade lifecycle is still in native QA.
One Pro license now activates on up to 5 computers in any mix of macOS, Windows, and Linux.
v6.2.122026-09-16
CodexUse now starts in light mode unless you have chosen dark mode. The Chrome Companion uses the same light design and matches the app version.
The Chrome Companion reconnects after app restarts and applies updated extension files when its tasks are idle. Version differences no longer trigger repeated reload requests.
ChatGPT Web setup now has three focused steps, copyable folder paths and connector settings, and links that open in Chrome. Entered values survive Back navigation, view changes, and retries; the form clears the API key after account creation or cancellation.
Updated workspace dependencies and fixed the dependency updater to include version overrides and avoid downgrading newer releases.
v6.2.112026-09-16
ChatGPT Web setup now uses only the bundled unpacked Chrome Companion. CodexUse opens Chrome Extensions, reveals the bundled folder, and copies its path without a Chrome Web Store listing or separate download.
The companion now stays at one stable per-user path and updates transactionally from each signed app bundle, so Chrome keeps the installation across CodexUse upgrades and interrupted updates recover safely.
Codex Low and Ultra routes now select ChatGPT Instant and Pro correctly, including the helper's normalized Max form, and package validation rejects a release that loses these browser mappings.
Automated browser tasks remain in Temporary Chat, while the companion popup reports its local connection and version for setup diagnostics.
v6.2.102026-09-16
ChatGPT Web now runs through the bundled CodexUse Chrome Companion in the user's regular signed-in Chrome profile. It no longer reads Chrome app data, uses remote debugging, launches a private browser, or sends AppleEvents to Chrome.
Setup shows the packaged companion folder and verifies the extension connection before checking the ChatGPT session, models, connector, tunnel, and one real tool-backed turn.
Instant now selects ChatGPT's native Instant reasoning mode, runtime status refreshes after an account opens, and tool turns no longer collide with other browser extensions.
Package validation requires the companion and authenticated loopback bridge and rejects builds that bring back the removed Electron browser helper or browser archive.
v6.2.92026-09-15
ChatGPT Web now connects to the ChatGPT session in your regular Google Chrome profile through Chrome's approved remote debugging path, so existing sign-in, passkeys, cookies, and network settings remain available. CodexUse creates only its own background task tabs and leaves Chrome signed in when disconnected.
The isolated Electron browser, throwaway Chrome sign-in transfer, cookie import, and packaged browser archive have been removed. A smaller sidecar now reconnects to Chrome, and upgrades terminate the retired sidecar before starting the new runtime.
v6.2.82026-09-15
ChatGPT Web setup no longer deadlocks when account settings load during a usage refresh, and a slow desktop request can no longer restart a responsive backend and erase the active setup.
If the backend does restart during setup, the open wizard now resumes the same setup automatically while keeping the tunnel key only in memory.
v6.2.72026-09-15
ChatGPT Web setup now brings its dedicated isolated Chrome sign-in window to the front through a signed macOS helper without requesting Accessibility, Automation, Passwords, or Keychain access.
v6.2.62026-09-15
Fresh installs now use the signed-in Codex account's native model catalog during ChatGPT Web setup, so setup can start before that account has opened its first managed Codex window.
Homebrew zap now removes CodexUse profile homes, browser data, caches, cookies, HTTP storage, and the legacy gateway launch agent along with app state.
v6.2.52026-09-15
ChatGPT Web setup now preserves the How it works confirmation across view changes, resets abandoned idle wizards before a new account, and returns to the confirmation step instead of failing on Connection.
v6.2.42026-09-14
ChatGPT Web connection checks now retry a direct model answer before changing a working account to Needs repair, and recovery guidance no longer asks users to change connector permissions.
v6.2.32026-09-14
Account details now remain visible for OpenAI, custom-model, and ChatGPT Web accounts while provider data loads or fails, with a direct retry instead of an empty pane.
ChatGPT Web setup now provisions its private model catalog before connector verification and waits for helper cleanup before rollback, preventing setup failures and abandoned tunnel runtimes.
Google, passkey, and Touch ID sign-in now runs in isolated Chrome data without reading browser data or requesting Passwords or Keychain access. Both browser paths deny page permission prompts and notifications, and the private browser removes unused microphone, camera, and Bluetooth declarations.
CodexUse no longer invokes macOS Passwords or Keychain for provider keys, Telegram tokens, Cloud Sync passphrases, trial tokens, browser cookies, or internal account capabilities. Saved secrets are encrypted in the app's owner-only local database instead.
Signed macOS builds refresh the embedded helper integrity manifest after final signing, then reseal, notarize, and verify the app so the shipped helper passes its compatibility check.
Existing installations migrate account state into a versioned namespace and repair older ChatGPT Web account records without allowing an older app process to erase newer account fields.
v6.2.22026-09-14
ChatGPT Web accounts can now answer Codex tasks and Work runs through a private browser session, with their own Codex home and browser-backed model routes.
A guided setup verifies the OpenAI tunnel and CodexUse connector end to end, supports password or Chrome passkey sign-in, and reports each setup, repair, and sign-out stage.
Browser-backed tasks now preserve follow-up context and Codex compaction across runtime restarts, contain cancellation and window failures, and finish interrupted streams with a clear reason.
ChatGPT cookies, prompts, responses, capability data, and runtime credentials stay out of Cloud Sync, exports, analytics, and diagnostic logs.
The website adds ChatGPT Web setup and policy guidance, troubleshooting, product screenshots, and updated search metadata for the new account type.
v6.2.12026-09-07
Work uses separate planner and worker accounts, isolated worker worktrees, and review before changes are applied to your working folder. Accept does not commit or push changes.
Accounts, onboarding, upgrade guidance, and CLI help now use consistent account and window wording. Refill controls describe background requests without promising extra quota or faster refills.
The website leads with Work, adds current Work and Accounts guides and screenshots, preserves redirects from retired docs, and improves mobile layout and accessibility.
CLI and Homebrew descriptions now match the app. Trial, license, provider limits, and website purchase activation instructions are aligned across public surfaces.
v6.2.02026-09-03
The website now checks unknown Mac architectures before offering the desktop installer. Saved-account, headroom, provider, trial, and purchase guidance describes the current product and its limits.
Welcome and Accounts guidance now uses verified account and window results. Your trial in facts keeps the first verified workflow on this Mac, including when optional analytics is off.
App checkout references now survive restart. Payment checks pause after 20 minutes, remain recoverable for 24 hours, and keep receipt-key activation available without another purchase.
License activation now rejects stale verification writes and separates existing-license recovery from new checkout completion. Verified payment records deduplicate sale identities and retain refund and revocation state.
v6.1.12026-08-31
Codex App account actions now distinguish shared-history switches from isolated windows, with direct account-row controls and a visible reset action for managed windows.
Shared-history mode now keeps the default Codex home and thread database while isolating account-specific renderer state, so existing threads remain available without stale account identity.
Codex window switching now verifies the source process, window topology, target account, and final activation before reporting success, and gives actionable guidance when unmanaged windows conflict.
v6.1.02026-08-28
The desktop now uses two primary workspaces: Accounts and Settings. Auto-roll controls live with accounts, connection tools share one settings rail, and first-run setup is one focused screen.
Account rows and details now use one readiness model for live, paused, locked, and recovery states, with clearer quota, search, activity, and next-action guidance.
Custom-provider setup now has a guided Add custom model flow that creates ready profiles transactionally and rolls back rejected provider or profile changes.
Cloud Sync now protects profile homes and configuration with transactional snapshots, crash journals, bounded cleanup, and exact rollback after interrupted or failed pulls.
Desktop support diagnostics now use shared key-aware redaction, while startup refreshes, profile switching, Codex window focus, Account Pool routing, and license verification report failures without corrupting saved state.
v6.0.62026-08-24
Custom-provider model setup now combines provider model lists with models.dev metadata. Selecting a model can fill its route, display name, context limits, reasoning options, image and tool support, and output-token format.
Existing custom-provider routes now refresh provider-owned metadata without overwriting saved endpoints or explicit user, probe, and runtime capability decisions.
The Model Providers editor now shows the context window, usable Codex context, and automatic compaction limit, and late discovery results cannot change another provider's draft.
v6.0.52026-08-24
Custom-provider traffic now runs inside the desktop RPC runtime instead of a separate gateway service. Upgrades drain and remove the legacy service with rollback protection, and app quit or update waits for active provider requests.
Model compatibility now uses an explicit per-capability plan. CodexUse can translate namespaced tools, serialize tool calls, preserve provider-default reasoning, and reject unsupported request shapes before they reach a provider.
Provider checks now use versioned request and response fixtures plus protocol fingerprints, so Codex App protocol drift and live provider rejections are visible and testable.
Custom-provider setup now shows transport fallbacks and thinking-effort choices, blocks profiles whose selected model routes are not ready, and keeps model settings repairable after a runtime rejection.
The public site, docs, pricing, release history, and download surfaces now use a clearer responsive layout with tighter platform and provider limits.
v6.0.42026-08-23
New custom-provider models now enable Apply patch edits and Parallel tool calls by default. Save and test still downgrades apply-patch support when the provider probe fails.
Capability and request-shape controls can now be corrected on an already-bound model after its running profile is stopped, without unbinding the model.
Provider compatibility errors now name the exact Model Providers toggle required for images, patch edits, namespaced tools, and parallel tool calls.
v6.0.32026-08-23
Saved accounts, license state, and preferences now reopen safely when SQLite sidecar files are missing. Healthy data stays intact, while damaged storage is quarantined and recovered without blocking desktop startup.
The desktop shell now opens immediately during app-state and profile bootstrap. Account loading, storage warnings, and retry actions stay inside the affected area, so Settings and other healthy areas remain available.
Storage recovery is coordinated across app processes, Account Pool revalidates after recovery, and saved-data writes report persistence failures instead of silently replacing durable state.
Custom-provider requests no longer forward parallel tool-call settings when the selected model capability is disabled.
Shipped features are easier to reach from Rotation and Settings, with corrected Model Providers help targets, aligned provider fields, and matching public documentation.
v6.0.22026-08-23
Kept CodexUse usable when account data cannot load. Account failures now show a retryable banner while Settings, Connect, Rotation, and other app areas remain available.
Isolated damaged profiles and startup services so one bad profile, Account Pool storage issue, local port conflict, license check, or usage refresh cannot terminate startup or hide healthy profiles.
Fixed a renderer crash caused by incomplete profile bootstrap data.
v6.0.12026-08-23
Fixed a 6.0.0 startup failure after upgrades when current and legacy profile homes coexist. CodexUse now preserves current runtime data, safely imports newer matching credentials, and keeps other profiles available if one profile home cannot be resolved.
v6.0.02026-08-23
Custom-provider profiles can now run OpenAI-compatible models from OpenRouter, DeepSeek, Groq, or a local endpoint inside the official Codex App. Each profile keeps an isolated Codex home, model catalog, reasoning-effort choices, and Keychain-backed credential.
The new native desktop gateway supports Responses and Chat Completions providers, function tools, reusable reasoning state, provider diagnostics, bounded concurrency, and persistent recovery across desktop updates.
Managed Codex profile launches, live switches, watchdog recovery, profile-home repair, and provider shutdowns now isolate failures and verify process identity before changing a running app.
The desktop and CLI now present OpenAI accounts and custom-provider profiles through one profile workflow, with clearer account, rotation, connection, and model-provider controls.
The public site adds provider setup documentation, focused provider guides, refreshed download and pricing paths, and generated route metadata for the expanded content library.
v5.6.32026-08-04
CodexUse now keeps the desktop window open when optional startup services such as the tray, updater, analytics, or embedded RPC runtime are unavailable.
Native window, updater, single-instance, child-process, socket, and background callback failures are now contained and logged instead of terminating the app.
v5.6.22026-08-04
CodexUse now opens even when Account Pool storage or one profile-auth migration is damaged. Affected data is preserved, unsafe profile-home conflicts stay blocked, and recoverable problems are shown without terminating desktop startup.
Profile credential reads now share cross-process locks while writes and crash recovery remain exclusive, preventing normal desktop and background refresh work from causing transient busy-profile and rate-limit failures.
v5.6.12026-08-04
Fixed a 5.6.0 startup crash when private profile-auth migration found an already-created destination. Matching stores now recover idempotently, while unverified destinations are deferred instead of overwritten or aborting app startup.
v5.6.02026-08-04
ChatGPT/Codex account switches now use one coordinated workflow across desktop, tray, CLI, and Auto-roll. Managed idle windows can switch accounts in place, while unsupported paths fall back to a verified relaunch.
Account switching is safer around active work and unmanaged windows. Generating turns defer through completion events, explicit or saved machine-local consent gates takeover, and process-identity checks prevent stale actions from interrupting the wrong app.
The CLI now supports private workspace-to-profile bindings, isolated profile runs, strict override handling, and default synchronization with an already-running ChatGPT/Codex app without launching a closed app unless requested.
Profiles and Account Pool now use stricter file coordination and per-process runtime leases, fail closed on damaged state, finish provider shutdowns cleanly, and no longer persist plaintext pool API keys.
The public site adds direct Pro purchase paths, a visible FAQ aligned with structured data, refreshed product guidance, and tighter prerendered routing and search metadata.
v5.5.12026-07-17
Paid Auto-roll, Account Pool, Telegram, and Cloud Sync screens now stay fully paused while locked, preserving saved settings without mounting protected queries or controls.
Reworked the Pro checkout and activation flow with clearer plan choices, automatic post-payment unlock, immediate entitlement refresh, and a direct receipt-key fallback when confirmation takes too long.
Refined Accounts and Rotation status views so paused, excluded, pending, and live account state is easier to distinguish at a glance.
Moved the desktop runtime to the published Effect 4 beta and hardened workspace dependency updates so prerelease packages stay on their intended release channel.
v5.5.02026-07-17
Rebuilt the desktop as a compact operations console: Rotation, Accounts, Connect, and Settings now expose live state and the next action without sidebar sprawl.
Added a dedicated Rotation command center with account ordering, a recovery lane, live activity, and safer arm and pause controls.
Redesigned Accounts, Account Pool, Telegram, Cloud Sync, CLI, and Support as focused workflows, including live Account Pool request visibility.
Reworked the landing site, docs, and blog with route-correct prerendering, stable hydration, clearer product proof, and preserved download analytics.
Hardened licensing and activation against forged seller and refund webhooks, duplicate first-open analytics, and native runtime mismatches during API deploys.
Bundled Bun 1.3.14 with the desktop runtime for consistent local execution across supported Macs.
v5.4.12026-07-17
The cockpit, accounts, Auto-roll, Account Pool, and settings now lead with live status and next actions instead of repeated configuration summaries, making the core multi-account workflow faster to scan and trust.
Auto-roll account participation now stays saved while the global policy is paused, with clearer paused labels, priority controls, and remaining-quota meters across account cards.
Telegram, Cloud Sync, CLI, updates, and support surfaces now use stable loading, empty, and failure states. Telegram also follows the app's canonical license state so navigation no longer flashes a false Free or Locked screen.
Conversion telemetry now distinguishes first opens, download intent, checkout sessions, activation outcomes, campaign attribution, and failure status without sending raw checkout session IDs.
v5.4.02026-07-13
Support both ChatGPT.app and Codex.app with safer account takeover, nested app-server detection, and bounded failed-launch cleanup.
Exclude individual accounts from Auto-roll and reset-window activation from either the account card or Auto-roll settings.
Add GPT-5.6 Sol, Terra, and Luna model aliases plus max reasoning effort across desktop and Account Pool.
Refresh the cockpit, account cards, Account Pool, and settings with denser status views and clearer primary actions.
v5.3.02026-06-28
The desktop dashboard got a broad UI refresh across Cockpit, Profiles, Settings, banners, modals, and update toasts, using one tighter component system with denser, scan-friendly layouts.
Cockpit and Settings now surface more real status in the first viewport, including clearer account rails, Account Pool setup, Telegram connection state, Cloud Sync state, CLI controls, and support actions.
The public site now ships as a React landing/docs/blog experience with refreshed hero, pricing, release, FAQ, and docs surfaces plus a static-host fallback for deep links.
Codex App control paths are safer around stale launches, reset-window model handling, and hash-route shell state so visible desktop chrome tracks the active route more reliably.
v5.2.92026-06-25
Codex App windows now keep native titlebar dragging while preserving the tinted sidebar and titlebar look.
Cockpit now uses a wider two-column dashboard with a pinned active-account status rail, and Configure panels scroll into view when opened.
v5.2.82026-06-25
You can now save your own OpenAI-compatible model providers in CodexUse. Add a provider that speaks Responses or Chat Completions, choose which one is active, and run a connection check before you rely on it.
Codex App launches and live switches now retry transient app-server verification failures, block duplicate launch clicks while startup is settling, and reserve Full Disk Access guidance for real permission blocks.
The desktop and landing UI got a native macOS polish pass: inset titlebar, translucent sidebar/titlebar surfaces, tighter Account Pool panels, and refreshed design-system tokens.
v5.2.72026-06-23
Codex App launch and close-after-switch recovery are safer: stale app-server children without a real main window are cleaned up before relaunch, and managed window detection no longer treats an orphan bridge as a running app.
Codex App discovery is narrower and quieter, avoiding broad application-folder scans while still checking the standard Codex.app locations.
Cloud Sync error messages now stay readable when a remote gateway returns HTML, and the shipped dependency set includes the latest small runtime and lint package updates.
v5.2.62026-06-21
Codex App takeover is safer for user-opened windows: explicit Take control now verifies the replacement bridge before closing the old window, and restores a plain shared window if recovery cannot finish.
Background auto-control no longer relaunches plain Codex windows, and reset-window activation now preserves only safe profile model settings instead of inventing fallback model IDs.
v5.2.52026-06-20
Reset-window activation now starts throwaway Codex homes with only the required model setting, avoiding user config hooks that could make the app-server exit before RPC startup.
Codex App recovery is safer around visible windows: passive health sweeps no longer kill bridge-missing launches, while direct launch failures still own ghost cleanup and now log the underlying technical cause.
v5.2.42026-06-20
Account Pool enable switches now respond immediately and stay in sync with the saved setting, even while the desktop runtime status refresh is still catching up.
v5.2.22026-06-20
Account Pool now has an explicit enable switch. When disabled, local OpenAI-compatible endpoints refuse requests so other apps cannot spend pooled accounts, while existing users with keys, sessions, or selected pools are migrated on.
Auto-roll can now start an exhausted account's 5-hour reset window right after switching away, and its watchdog checks faster when the active account is close to the switch threshold.
Reset-window activation now pins throwaway Codex runs to a supported model, and the default Codex presets use gpt-5.5 so ChatGPT-account auth avoids blocked *-codex model IDs.
v5.2.12026-06-19
Codex App automatic control is now opt-in and backed by a one-time settings migration, so existing windows are not taken over unless the setting is explicitly enabled.
Reset-window activation now skips accounts with known exhausted quota, and the cockpit shows a clear no-op warning when every launchable profile is still rate-limited.
v5.2.02026-06-19
Account Pool now reports a real desktop loopback endpoint, keeps integration snippets compact, and routes parallel requests away from already-busy sessions before considering rate-limit percentages.
Telegram remote control is cleaner end to end: bot tokens live in Keychain, setup shows the exact /link next step, phone commands stay aligned with the menu, and remote replies include the latest progress, tool output, diffs, and local image artifacts.
Codex App launch and recovery paths are stricter. Failed launches clean up ghost main processes, Open Codex stays on the shared-home path, stale Telegram targets fail explicitly, and app-server verification gates reported success.
Settings and Cloud Sync got a pass for clarity: saved secrets show explicit saved-state notes, diagnostic action grids have room to breathe, and sync snapshots no longer carry local bookkeeping that can make a good pull look stale.
v5.1.82026-06-18
Telegram bridge now runs real bot polling from the desktop runtime. Allowed chats can use /status, /profiles, /switch, /codex_apps, /id, and /help while trial or Pro access is active.
Reset-window activation is sturdier for large account pools. CodexUse now activates accounts one at a time with visible per-account progress and a longer timeout for the slow activate-profiles path.
Codex App live switches now verify the app-server child and bridge health before reporting success, retry recoverable launch failures, and log enough evidence to diagnose Auto-roll switches that land on the wrong window.
v5.1.72026-06-18
The public landing page now leads with a tighter cockpit-first story: live account headroom, Auto-roll context, copyable Homebrew install, clearer pricing, stronger FAQ/release layout, and better mobile spacing.
Codex App control is sturdier around reset windows. Pro users can activate reset windows from the cockpit, and recoverable live-switch launch failures get one quick retry before reporting failure.
Support and runtime recovery are sharper. Settings can reveal the exact support.log file, desktop RPC timeouts now kill the stuck child so it restarts cleanly, and Auto-roll confirmation failures queue a single delayed retry instead of stacking retries.
v5.1.62026-06-17
Rate-limit telemetry is now app-server-only. OAuth usage estimates, passive rollout-file reads, and stored profile snapshots no longer supply quota truth for the dashboard, CLI, Auto-roll, or account choice.
Auto-roll and CLI best-profile runs now stop on unknown live usage instead of choosing from stale cache. The current account is confirmed through the Codex app-server before switching.
Rate-limit cache seeding now accepts only Codex RPC snapshots, so rejected OAuth or passive seeds cannot block a live refresh or keep old usage on screen.
v5.1.52026-06-17
Rate-limit telemetry now uses confirmed Codex RPC snapshots only. OAuth, passive, and profile-cache estimates no longer paint the dashboard or drive Auto-roll decisions.
Codex App controls are sturdier with current Official Codex builds. Plugin install unlock now also keeps the required config flags enabled, model-list patches stay scoped to model surfaces, and injected controls work correctly in draggable titlebar areas.
Codex App handoff and export can now find thread metadata in the newer sqlite database layout as well as the legacy state file, including automation-run records when available.
The dashboard cockpit, account roster, and upgrade flow are calmer: account groups show stale or sign-in status more clearly, the Codex App panel is unified around the account list, and only one blocking modal mounts at a time.
v5.1.12026-06-17
Every install now starts with a 7-day CodexUse Pro trial. The trial grant is issued by the CodexUse API, cached with signed local state, and can keep working through short offline windows without turning local files into the source of truth.
Trial access now gates the real execution paths, not just the visible UI. Saved profiles, CLI/RPC actions, Auto-roll follow-ups, Account Pool, Cloud Sync, and Codex App controls all re-check trial or Pro entitlement before using paid features.
The app and landing site now lead with the trial experience: first-run Pro welcome, trial countdown, trial-ending reminders, download-first CTAs, and clearer upgrade copy when the trial ends.
v5.1.02026-06-15
Auto-roll now requires fresher confirmed usage before switching accounts. When quota telemetry is stale or comes from an OAuth estimate, CodexUse confirms the active account through the Codex RPC path before taking the switch.
Manual usage refreshes can accept real reset jumps while still confirming material telemetry changes, so freshly reset accounts update without letting optimistic reads silently replace trusted low-headroom data.
Account cards are calmer when supplemental quota windows are present. Core limits stay prominent, extra limits collapse under a compact row, and supplemental windows no longer trigger critical styling for the main account state.
v5.0.92026-06-15
Rate-limit snapshots now persist across desktop reloads and are pruned when profiles are deleted or replaced through Cloud Sync, keeping low-headroom safeguards intact without retaining removed-account telemetry.
Auto-roll priority is clearer for mixed ranked and unranked accounts. Optional ranks are still tried first, while unranked eligible accounts now sort by the most remaining usage.
Cloud Sync remote snapshot deletion now uses the app confirmation dialog, so the destructive action matches the rest of Settings instead of relying on a browser confirm prompt.
v5.0.82026-06-15
The current Pro offer now uses the SUMMER50 checkout code across the app, CLI license status, and public landing site while keeping the 50% lifetime discount.
v5.0.72026-06-15
Cloud Sync now encrypts snapshots before upload with a local passphrase. Pushes store encrypted data only, pulls can migrate legacy plaintext snapshots, and the app can remember the passphrase in Keychain for the current license.
Cloud Sync controls are clearer and safer across the app and CLI. Passphrase input is explicit, remote snapshot status distinguishes encrypted and legacy snapshots, and users can delete the remote cloud snapshot when they need a clean reset.
The desktop RPC runtime now runs as a supervised child process with startup timeouts, request timeouts, heartbeat checks, crash logging, and restart backoff. CodexUse can recover the account/status runtime without taking down the whole desktop app.
macOS stability diagnostics are stronger. Support info can include redacted native crash summaries, and the release build now prepares and signs the Electrobun CLI before packaging so local release runs fail earlier and more clearly.
v5.0.62026-06-14
Codex App handoff is stronger for running threads. Account switches can capture active Codex App threads, resume each thread through the app-server, send the continue message in the same conversation, and retry failed sends instead of dropping activation state.
Rate-limit checks now have a direct Codex OAuth usage path. CodexUse can refresh saved Codex OAuth tokens, read usage from the Codex backend, and show supplemental quota windows such as Codex Spark 5-hour and weekly limits beside the core windows.
Settings saves are safer when multiple controls update close together. Runtime settings patches now build from the latest settings state so Account Pool, Codex App controls, and queued saves do not overwrite each other with stale full-object writes.
Codex App process control handles partial exits better. If the main Codex process disappears but the app-server child is still alive, CodexUse keeps that child tracked and can stop it cleanly instead of marking the window stopped too early.
The dashboard quota and settings UI is tighter for dense account lists, with compact supplemental limit rows, clearer status badges, and a React Doctor policy update for the new settings mutation pattern.
v5.0.52026-06-14
Settings > Support now has a visible app update card. CodexUse can check for desktop updates, download a ready build, show the installed and available versions, and restart to install from the same support surface.
Codex App account switches are safer while a response is still generating. Auto-roll and manual cockpit switches now defer the Codex App follow-up until the source response finishes, keep the activation debt alive, and tell you when the target window will update later.
Codex App restarts and stops now wait for the whole process tree, including the app-server child, before relaunching or marking a window stopped. That prevents relaunches from reopening into partial or empty state after the main process exits first.
The dashboard and public site are cleaner for the 5.x line: remaining-limit chips mute the non-depleted window when another quota window is exhausted, macOS menu shortcuts show the expected Command-Q and Command-W accelerators, and landing/docs copy now consistently presents CodexUse as the Codex control plane.
v5.0.22026-06-14
CodexUse now has a single cockpit for Codex App accounts, Codex App controls, Auto-roll policy, profiles, and CLI status. The old split settings surfaces now route into one account-first view so setup and switching state stay together.
Auto-roll prompts are safer. CodexUse owns each switch decision, shows the countdown only after a visible dashboard or Codex App overlay acknowledgement, cancels stale decisions when you stay or switch manually, and keeps draft handoff attached to the accepted switch.
Codex App account truth is more reliable during manual switches and managed-window launches. Managed windows now feed the same sync status as observed windows, switching state clears after the real target is seen, and Settings no longer needs a separate repair action to show the current account.
Rate-limit telemetry is stricter around reset windows. Recent low-headroom snapshots are no longer overwritten by optimistic fresh 100% reads before a real reset boundary, refresh cooldown state is exposed to the UI, and sign-in failure copy points users to the account fix.
v5.0.12026-06-13
CodexUse no longer asks macOS for broad automation or filesystem access when finding and activating Codex App. App discovery now stays on known Codex.app paths, profile focus uses the safe macOS open path, and the UI reports activation requests honestly when focus cannot be verified without TCC prompts.
Desktop startup no longer launches the user's login shell just to learn PATH. CodexUse now builds a narrow macOS tool path from known developer directories, avoiding shell rc side effects that could trigger protected-folder prompts.
CLI install and update checks now find common nvm, fnm, Volta, asdf, Bun, Homebrew, Cargo, and CODEX_PATH_HINTS locations without sourcing shell config files.
v5.0.02026-06-12
CodexUse startup and app-state writes are much faster. Large profile maps no longer pay repeated JSON clone costs, deleted profile keys stay deleted, and first paint avoids blocking Codex App discovery.
Codex App in-app controls are more reliable on cold start. The overlay now injects at document start, restores handoff and Auto-roll targets from saved settings, and keeps bridge requests visible in backend logs for support.
Account Pool local API latency now stays local. Hot-path auth avoids durable store reads and online license refresh per request, with debounced usage writes and stale license status used safely between refreshes.
Rate-limit refresh no longer floods local Codex app-server workers. Account probes are capped, rollout quota reads stay scoped to the current profile, and tray refresh churn is reduced.
The dashboard and landing surfaces were polished for 5.0.0: settings styles were split into tighter token files, new hero screenshots ship in WebP, and update/download states use clearer CodexUse copy.
v4.0.02026-06-02
Pro checkout can now move from purchase to activation without copy/paste. CodexUse creates a checkout session, watches for the matching Gumroad receipt, verifies the license locally, and keeps an honest waiting state with manual activation fallback.
The new activation API stores Gumroad Ping licenses by checkout session, serves each license key once to the desktop verifier, prunes old activation records, and can forward non-CodexUse Gumroad payloads when a fallback endpoint is configured.
First-load recovery is safer when local app-state storage or profile bootstrap stalls. CodexUse now bounds app-state initialization, repairs only the damaged app-state document, and skips nonessential Codex App enrichment instead of leaving the app on Loading.
Settings > Support now includes stronger one-click diagnostics with customer-copyable PostHog distinct/session ids, diagnostics bundle ids, update and license state, bootstrap context, and redacted recent logs.
Support analytics now has dedicated bootstrap and customer triage dashboards for stuck Loading sessions, desktop RPC phases, backend bootstrap phases, diagnostics bundles, customer errors, and license activation funnels.
The shipped dependency set is refreshed and trimmed, including Base UI 1.5.0, TanStack Query 5.100.14, TanStack Router 1.170.10, lucide-react 1.17.0, and TypeScript ESLint 8.60.1.
v3.9.92026-05-27
Codex Apps can now activate reset windows for saved profiles with one confirmed action. CodexUse sends a minimal Codex request for each launchable profile, refreshes rate-limit data afterward, and shows completed, skipped, or failed activation results per account.
Codex App profile switching is safer when restarts overlap. CodexUse now treats the latest requested target as the owner, cancels stale restart debt, and avoids an old retry relaunching the wrong account after you manually switch profiles.
The Codex Apps sync card now merges live window observations with saved restart state, tracks reset-window activation events, and can open the event log directly for faster diagnosis of stale or mismatched app-account state.
Reset-window activation uses an isolated temporary Codex home and only writes refreshed auth back when the source auth has not changed in flight, preventing stale token data from overwriting a newer profile switch.
v3.9.82026-05-26
Codex Apps now has its own dashboard surface for managing OpenAI Codex app windows, with clearer install state, managed-window counts, main-app status, profile rows, and focused launch/stop/restart controls.
OpenAI Codex app sync is easier to trust. CodexUse now records observed profile names, hashed identity matches, match source, and recent activity so stale, synced, and unknown app-account states are easier to diagnose.
Auto-roll is more resilient when quota snapshots go stale or Codex App window proof is delayed. A server watchdog refreshes old usage data, tracks one decision id across eval, switch, auth verification, target activation, and observed app-account match, and preserves pending activation state until the target proof is ready.
The CLI profile switch and auto-roll paths now use the same profile-targeted OpenAI Codex app activation logic as the desktop runtime, including unmanaged-window checks, scoped launch/focus behavior, and recorded outcomes.
Settings now has a dedicated Support page for docs, email support, app updates, license state, CLI version, support-log location, support-folder opening, and a stronger copy-support-info fallback when browser clipboard access is denied.
The shipped dependency set is refreshed, including TanStack Query/Router/Virtual patch updates, Hono 4.12.23, @hono/node-server 2.0.4, and TypeScript ESLint 8.60.0.
v3.9.72026-05-23
Codex Apps controls now follow the same visible-profile rules as the rest of CodexUse. Active trials can orchestrate unlimited profile windows, and Pro keeps them unlocked after the trial.
Codex App window actions are cleaner and more consistent. Launch, focus, stop, restart, launch-all, and stop-all now filter unavailable profiles through one shared access path and return clearer upgrade guidance when access ends.
Live rate-limit refresh now uses the same license-visible profile list, so unavailable profiles no longer leak into background headroom probes while active-trial and Pro profiles still get usage monitoring.
The CLI now enforces the Pro requirement for `codexuse profile autoroll`, and the CLI help/README call that out directly.
Pricing, docs, feature copy, and generated landing pages now describe CodexUse as a local Codex App orchestrator plus CLI profile manager, with shipped process orchestration and draft handoff separated from full session cloning.
v3.9.52026-05-22
Codex Apps can now orchestrate Codex App windows per saved profile from Settings. Launch, focus, restart, or stop a profile window directly, and use bulk launch/stop controls when multiple profiles are available.
Already-open Codex App windows are easier to control. CodexUse now detects running Codex app processes, maps matching windows back to saved profiles, and shows PID, app-server PID, running state, and uptime in one live inventory.
Codex App profile launches are safer and more verifiable. Profile actions prepare an isolated CODEX_HOME, wait for the matching app-server process, and only mark the window running after the expected profile instance is observed.
The dedicated Codex Apps Settings tab is wired through the desktop runtime and native client, so the renderer no longer has to guess at profile-window state.
The public site now explains Codex App multi-account orchestration more clearly, with a new Codex Apps docs page, a new setup guide, updated sitemap/llms metadata, and safer JSON-LD generation for generated docs and blog pages.
v3.9.42026-05-22
Codex Apps is now a dedicated Settings surface for Codex App windows. It detects profile-matched Codex app processes, shows running state, PID, app-server PID, and uptime, and lets you orchestrate focus, stop, restart, or launch for profile windows from CodexUse.
Auto-roll can now warn before a profile fully runs out: Settings adds a low-remaining usage alert with a configurable threshold, and CodexUse sends one desktop notification per quota window when a profile drops below it.
Codex App activation after auto-roll is more direct and verifiable. CodexUse now focuses or launches the target profile window, logs the activation path, and records a fresh sync only after the target proof is verified.
Desktop reopen recovery is more resilient when a stale single-instance lock points at a stuck CodexUse process tree. Takeover now verifies the owner, terminates descendants, retries acquisition, and lets the user retry from the startup dialog if needed.
The shipped frontend dependency set is refreshed, including Vite 8.0.14, TanStack Router/Virtual patch updates, and Framer Motion 12.40.0.
v3.9.32026-05-20
Desktop window focus now uses Electrobun's `activate()` path directly for main and in-app browser windows, avoiding the deprecated `focus()` bridge and keeping show behavior aligned with current Electrobun releases.