Changelog

v2.9.3 July 29, 2026 Latest

Stability update

• Long chats no longer freeze the app. Two causes fixed: restoring a heavy conversation rendered oversized messages on the interface thread (measured 1–3 seconds of freeze per pair of messages), and scrolling up kept loading older messages forever until hundreds of live widgets dragged everything down. Restore is now capped and lightweight; older history stays safely in the database. A built-in stall detector also now records any real interface freeze in the log with its exact duration, so future reports get diagnosed from evidence.
• The terminal behaves like a real terminal. Ctrl+C now stops a running server (with text selected it copies, like VS Code); Ctrl+V pastes instead of crashing the app — the paste handler simply didn't exist.
• The AI can no longer claim work it didn't do. New guards catch "Fixed all 70 lines" and "Pushed, commit abc1234" style claims made without a single tool call — both caught in real logs with invented results — and force the actual work. A finished task must also state what it actually did: a model that wraps up inside its hidden reasoning and leaves the reply empty is now made to write the summary.
• You can see what the AI is doing. The status bar shows which skill is steering the current task, live, and clears when it finishes. The Tasks panel now marks the task being worked with a spinner instead of showing idle circles through minutes of activity. The "Compacting conversation" overlay always closes — it previously had no off switch on the compaction path and could spin forever.
• Every skill is now viewable and editable. Click any of the 251 skills to read its complete SKILL.md — rendered or raw source — and edit it. Editing a built-in creates a personal copy that safely overrides it by name; delete the copy to restore the original. Edited copies keep their identity and still reach the original's reference and template files.
• MiMo with a pay-as-you-go key now sends images natively instead of taking a slow OCR detour through the server — lossless and much faster. Token-plan keys keep the OCR fallback that makes images work for them at all.

v2.9.2 July 28, 2026

251 skills, live skill indicator, and the freeze fixes

• 32 SEO skills added — 251 built in. Full search-optimisation coverage: site audits, technical SEO, content briefs, keyword clusters, backlinks, schema, sitemaps, hreflang, local & maps, e-commerce, programmatic pages — plus GEO/AEO skills for showing up in AI search (ChatGPT Search, Perplexity, Google AI Overviews). Just say "use the seo-audit skill".
• See which skill the AI is using — live. The status bar now shows ⚡ Skill: seo-audit the moment a skill loads, and clears when the task finishes. Hover it for the full list. If you mistype a skill name, nothing appears — so you know instantly instead of guessing. Fixed along the way: the indicator was only wired to one of the two chat modes; it now works in both.
• Skill search that actually finds skills. With 251 skills, the old search fell over: typing "3d website" found nothing because the skill is named 3d-website-architect. Spaces and hyphens now match each other, every extra word narrows the result, tags are searched, and the best match moves to the top — in the Skills window and the "/" menu alike, with one shared ranking.
• Long chats no longer freeze the app. The cause of the "aggressive freezing": scrolling up kept loading older messages forever — sessions reached 139 live messages, roughly 700 rich-text widgets in one column, and every streamed chunk re-laid-out all of them. The chat now caps at 80 live messages; older history stays safely in the database.
• Ctrl+V in the terminal no longer crashes Cortex. The terminal had no paste handler at all — pasting raised an uncaught error that left the app in a broken state. Paste, copy, cut and select-all all work now.
• Restart offer after switching projects. Opening a different project resets what you can see, but the AI's loaded context, search index and file watcher still point at the old project. A small corner card now offers a one-click restart — it waits until your chat history has finished loading, and only appears on a genuine switch, never on first open.
• Installer hardening. The 251 skills are data files, and a packaging slip could have shipped an installer with zero skills and no error. The loader now searches every location a build can put them and reports loudly if they are missing — with tests locking down the packaging itself.

v2.9.1 July 24, 2026

The Skills release — 219 expert working methods, built in

• 219 built-in skills. A skill is a page of professional working instructions the AI follows — how to debug properly, plan before coding, write tests first, review code, harden security, tune performance, design an API, and build beautiful frontends. Adapted from the published playbooks of industry leaders (Addy Osmani, Matt Pocock, Meng To, Emil Kowalski, Anthropic and others) and fitted to Cortex's tools. Every skill is a plain readable file you can open, edit, or replace.
• Engineering and design, both covered. Alongside the visual skills there is now a full engineering set: security hardening, performance optimisation, API and interface design, git workflow, CI/CD, observability, code simplification, and accessible frontend engineering.
• Skills cost you nothing until you use one. All skills start switched OFF. Just name one in your message — "use the apple-design skill" — and the AI loads it for that task only. Turn a toggle ON only for a skill you want applied to every message.
• The AI can also pick a skill by itself. Every conversation carries a compact list of names; when your task matches one, the AI loads it before starting. No setup, no clicking.
• Type "/" to call a skill. The chat box now has a slash menu — press "/", search, pick a skill, and Cortex writes the instruction for you.
• New cost banner in the Skills window. Switched-on skills are added to every message you send, which was invisible before. The Skills window now shows exactly what that costs — "27 skills on · about 43,900 tokens added to every message" — and turns amber, then red, as it grows. Each skill also shows its own size before you switch it on.
• Premium website building. Dedicated skills for 3D and animated websites: Three.js scenes, WebGL hero objects, GSAP scroll choreography, animated backgrounds and cinematic text reveals — with a self-critique step where the AI previews its own page and iterates before showing you.
• Anti-whack-a-mole debugging built in. The debugging skill enforces the discipline that actually kills bugs: stop, preserve evidence, reproduce, fix the root cause, add a regression test — never guess-and-patch.
• Redesigned Skills window. AI → Browse Skills is now a proper panel: card layout, real toggle switches, search, a detail pane, and a live enabled count. Your toggles survive restarting Cortex.
• Add your own skills. Drop any standard SKILL.md folder into ~/.cortex/skills (or a project's .cortex/skills) and it appears automatically — and can even override a built-in by using its name.

Fixed

• The AI no longer redoes work it had already done. The bug: after preparing a set of edits, the AI would re-read the file, see the older version still on disk, announce "the edits never landed" and redo the entire job — burning minutes and sometimes producing worse output the second time. It now sees its own prepared version when it reads back, so it trusts work that is genuinely there.
• A skill you switched on was being sent twice. The bug: naming a skill that was already toggled ON delivered its full text a second time — on one skill that was 14,750 characters of duplicated instructions, leaving the AI with two copies and no way to tell which one applied. Cortex now points at the copy already in place.
• The interface no longer freezes when you switch theme. The bug: changing between dark and light locked the window for several seconds on long chats. Visible parts are now restyled first, so the switch looks instant.
• Skill descriptions written in multi-line form are read correctly. Some skills lost their entire description to a formatting quirk, which made them invisible to search and to the "/" menu.
• A large skill library no longer inflates every message. With 219 skills installed the catalog alone was adding roughly 8,000 tokens per message. It is now under 1,100, whatever the library size.
• The "/" menu no longer stops at 8 skills, and picking one responds the moment you click rather than after you release.
• The "file changed outside the editor" pop-up no longer appears when you press Run without having edited anything — Cortex silently takes the newer version. Also fixed a crash when opening the Memory Manager.

v2.9.0 July 23, 2026

The reliability release — the AI can no longer claim work it didn't do

• Every edit is now verified against the real file. The bug: sometimes the AI said "done ✅" but the file was unchanged — a failed edit reported itself as success, and the AI honestly repeated the lie it was told. Cortex now re-reads the file after every edit and write; if the change is not really there, the AI gets an error and must retry. It can no longer report unverified work as complete.
• The AI uses proper edit tools again. The bug: the AI had drifted into writing throwaway patch scripts and running them through the terminal to change files — fragile edits that often silently did nothing. Cause found: internal safety rules blocked the proper Edit tool while always allowing the terminal, teaching the AI the wrong habit. The rules are fixed and patch-script editing is now explicitly forbidden.
• Permission cards no longer expire. The bug: if a permission card appeared and you stepped away — tea, a call, anything over 30 seconds — the task silently died with a false "user rejected" recorded, and clicking Allow afterwards did nothing. Cards now wait as long as you need, at zero cost (no background polling), and one click resumes the task instantly. Stop still reacts immediately, and genuinely stuck commands still time out.
• No more "please accept" after you already accepted. The bug: the AI kept asking you to approve edits you had approved minutes earlier — nobody ever told it your click happened. It is now notified the moment staged edits reach the disk.
• Your work survives a crash. The bug: edits the AI had prepared but not yet saved lived only in memory — a power cut or forced restart threw them away silently. They are now mirrored to disk the moment they are staged and recovered on the next launch.
• Ctrl+S can no longer erase the AI's edits. The bug: with a file open in the editor, if the AI changed one part and you edited another, saving replaced the AI's work with your older copy — invisibly. Cortex now detects that the file changed on disk and asks you: Save anyway, Reload from disk, or Cancel. Save All is covered too. And if you changed nothing yourself, there is no pop-up at all — Cortex simply takes the newer version silently (e.g. pressing Run right after the AI edited the open file). Nothing is silently thrown away in either direction.
• Explorer sees files created anywhere in the project. The bug: files the AI created inside subfolders never appeared until you clicked refresh — the file watcher only watched the top level. It now watches the whole project tree, and newly created folders are picked up automatically.
• Smoother chat during long answers. Long Plan/Ask explanations made the chat stutter — every redraw re-rendered the entire answer. Redraw now adapts to answer length: 63% less work on long answers, short replies unchanged.
• Much faster under load — background monitoring removed. Cortex was watching RAM and CPU constantly, and on busy machines it reacted by pausing the AI several seconds on every step, running cleanup passes, and writing emergency saves — the biggest hidden slowdown in the app. All of that is gone: system checks are now a once-per-hour heartbeat and nothing ever slows the AI based on them. A blocking measurement call (72 wasted seconds per hour) was also fixed, and leftover debug messages that filled half the log are gone.
• Cleaner installs. Cortex now installs into the correct 64-bit Program Files folder, and the installer verifies the download's checksum before running it.
• 110 automated tests added — the first full test suite, covering every bug fixed in this release so none of them can quietly return.

v2.8.9 July 20, 2026

Fixed

• Gray flash when the IDE opens. The bug: for a split second at startup, a bright gray rectangle flashed on screen before the dark loading screen appeared. The window now stays invisible until its first picture is fully drawn, so startup goes straight to the dark Cortex loading screen — no flash.
• Your edits no longer revert by themselves. The bug: sometimes text you (or the AI) had just fixed would quietly change back to the old version minutes later. Cause: edits the AI had queued earlier could overwrite newer changes when a task finished or the IDE closed. Outdated queued edits are now detected and discarded instead of applied.
• Git conflicts are now YOUR decision. The bug: when a git pull or rebase hit a conflict, the AI silently picked which version of the code to keep — sometimes resurrecting old text you had already replaced. The AI now stops, shows you both versions in plain language, and asks which one to keep before continuing.
• Empty gaps in chat removed. The bug: large blank spaces appeared between chat cards while the agent worked, and every short chat line carried extra empty space under it. Both causes are fixed — cards now sit tightly together.
• Spelling check works in the installed app. The bug: the red underline for misspelled words in the chat box worked during development but was dead in the installed version — the dictionaries were missing from the build. They're included now.
• Changed Files and Tasks rows no longer cut off. The bug: after the fix that kept these cards from covering the chat input, long lists showed rows sliced in half. The cards now size themselves to the window, so rows are always fully visible and long lists scroll cleanly.
• Restore and memory pop-ups fully readable. The bug: the "Restoring chat history" and "Saving to memory" cards clipped their text and spinner when the message was long. The cards now grow to fit their content.

v2.8.8 July 20, 2026

Fixed

• Chat text corruption fully fixed. The bug: words displayed merged ("port 8083" became "port8083") or split ("expected" became "expe cted"), commands in code cards lost words and spaces, cards showed wrong titles or came up empty — and the Copy button copied the broken text. The root cause was in how streamed text was reassembled; spaces and line breaks were being silently dropped. Text now displays exactly what the AI wrote, and Copy gives you the exact correct command.
• "Restart Now" actually restarts. The bug: clicking Restart after changing the theme closed the IDE but it never came back — you had to reopen it yourself. The restart now waits for the old instance to fully close, then relaunches automatically.
• Files created by the AI now appear in the Explorer. The bug: the file tree only updated after a manual refresh. It now refreshes automatically within seconds — including files the AI creates while working.
• Right-click menu readable in light mode. The bug: right-clicking a file in the Explorer opened a dark menu with dark text in light mode. The menu and the rename dialog now follow the active theme.
• Appearance settings readable in both modes. Theme card labels and hover text in Settings → Appearance now display correctly in dark and light mode.

v2.8.7 July 19, 2026

New

• Grok 4.5 model added. xAI's newest model — 500k context window, 32k output, vision support — is now available through OpenRouter.
• Settings shortcut. Settings now opens with Ctrl+, from the File menu — the shortcut was previously advertised but never worked.

Fixed

• WSL, CMD, and Git Bash terminals failed to launch. The bug: picking WSL as your terminal shell failed instantly with "Invalid command line argument: -NoLogo" — a PowerShell-only flag was being passed to every shell. Each shell now launches with its own correct arguments, and Git Bash loads your profile properly.
• Help menu items did nothing. The bug: most Help menu entries were dead — no action, or links to pages that don't exist. Every item now works: Documentation, What's New, Skills, MCP guide, Troubleshooting, and Send Feedback all open their real destinations. Leftover entries for features Cortex doesn't have were removed.
• Skills browser crashed on open. The bug: AI → Browse Skills and Help → Skills silently failed every time. The browser now opens, lists your skills with on/off checkboxes, and explains how to create your first skill if you have none. A second bug that broke the whole skills system when a skill file used list-style tags is also fixed.
• Sidebar icon names showing only once on hover. The bug: hovering Explorer, Search, AI Tools, Problems, or the Settings gear showed the name the first time but often not again. Tooltips now appear on every hover, reliably.
• Keyboard shortcuts reference cleaned up. The bug: the F1 shortcuts dialog listed shortcuts that don't exist (New Window, Quick Chat, Command Palette, Find in Files, Toggle Review Panel) and a few wrong ones. Every listed shortcut now actually works, wrong labels were corrected, and working shortcuts that were missing from the list were added.

v2.8.6 July 18, 2026

New

• Kimi K3 model added. Moonshot AI's latest flagship — 1M context window, vision support — is now available through OpenRouter. Pick it from the model dropdown like any other model.
• GPT-5.6 models now work with your own OpenAI key. The full GPT-5.6 family — Luna, Terra, Sol, and their Pro versions — was previously available only through OpenRouter. Now you can also use them directly with your OpenAI API key.

Fixed

• Chat input disappearing under long task lists. The bug: when the agent was working through many tasks, expanding the Tasks (or Changed Files) card above the chat input grew it taller than the window — and the input box was pushed completely off-screen. Both cards now scroll internally past a fixed height, so the input always stays visible.
• Ctrl+S silently doing nothing after the editor's right-click menu. The bug: after using Format Document / Format Selection from the context menu, pressing Ctrl+S didn't save — no error, nothing — until you closed and reopened the file. Saving now works no matter where focus is inside the editor.
• New files not appearing in the Explorer. The bug: files created by you or the AI didn't show up in the sidebar file tree until you clicked refresh. The tree now updates automatically about a second after any file is created, and catches up right after the AI finishes working. Newly created folders are watched too.
• Unreadable terminal text in light mode. The bug: most terminal output (PowerShell's default text color) appeared as pale gray on the light background. All terminal colors are now dark, high-contrast tones in light mode.
• Confusing OpenRouter error on connection failures. The bug: if the connection to OpenRouter failed while starting a chat, Cortex showed a cryptic internal error instead of retrying. Now it automatically retries, and if it still fails you see the real reason (timeout, connection refused, etc.).

v2.8.5 July 16, 2026

New

• Paste more image formats. WebP, AVIF, and SVG images can now be pasted into chat — previously only PNG and JPEG worked. SVG files are converted automatically so every AI model can read them.

Fixed

• Pasted images now actually reach the AI. The bug: you pasted an image, but Cortex silently removed it before sending the request — so the AI said it couldn't see any image, or described an older screenshot from earlier in the chat. This affected every model except MiMo (Claude, GPT, Grok, Gemini, DeepSeek — with your own key or via OpenRouter). Now the image is delivered with your message every time.
• Wrong "Subscription Required" message for vision models. The bug: pasting an image with a model that can see images natively (like Claude or GPT) wrongly showed the subscription wall, because Cortex sent every image to the subscription-based OCR service. Now images go straight to your selected model when it supports vision — OCR is only used for models that genuinely can't see images.
• Switching models then pasting an image used the old model. The bug: if you changed the model and immediately pasted an image, Cortex still checked the previous model's abilities and picked the wrong path. Now the switch takes effect instantly.
• The AI kept describing an old image. The bug: leftover images from earlier messages stayed attached, so a new paste could get answered with a description of the previous image. Old images are now cleared before each new message.
• DeepSeek chats failing with an HTTP 400 error. The bug: mid-conversation, DeepSeek could reject the request and the chat stopped with a "Bad Request" error. Fixed — long conversations with tool use now continue normally.
• Safety net when direct image sending fails. If sending an image straight to the model fails for any reason, Cortex now falls back to OCR text extraction instead of showing an error.

v2.8.4 July 15, 2026

Fixed

• "Restart Now" after a theme change now actually restarts. Clicking Restart Now closed the IDE but never brought it back — a new instance either launched incorrectly or was blocked by the still-closing previous one. The restart now waits for a clean handover and relaunches reliably.
• MCP server status no longer stuck on "connecting". The MCP Servers panel could show a server as connecting forever even though it had connected fine seconds later. The panel now keeps checking until the real status appears.

v2.8.3 July 14, 2026

Fixed

• In-app updates now download correctly. Clicking "Update Now" could fail with an "HTTP 403 Forbidden" error — our security layer was blocking the app's own download request. Updates now download and install as expected. (One-time note: if you're on 2.8.1 or 2.8.2, download 2.8.3 once from the website — after that, in-app updates work automatically.)

v2.8.2 July 14, 2026

New

• Claude Sonnet 5. Anthropic's newest Sonnet, available both natively (your Anthropic key) and via OpenRouter — 1M token context, 64K output, extended thinking supported.
• GPT-5.6 family via OpenRouter. Six new OpenAI models in a new "OpenRouter — OpenAI" group: Luna, Terra, and Sol, each with a Pro reasoning variant — all with 1M context and 64K output. Every one verified with live test calls before shipping.
• Antivirus & Performance guide. The docs now include one-minute exclusion steps for every major antivirus product, like other IDEs recommend.

Fixed

• Claude models via OpenRouter couldn't connect. The OpenRouter versions of Claude Opus 4.8/4.5, Sonnet 4.5, and Haiku 4.5 pointed at model ids that don't exist on OpenRouter (it names them with dots, not dashes). All four are corrected and confirmed working with real completions against OpenRouter's live catalog.
• Clearer provider errors. When a provider rejects a request (for example, a model that isn't offered in your region), Cortex now shows the provider's actual reason and what to do about it — instead of a bare "HTTP 403".
• Removed Grok 4.5. xAI doesn't offer it in many regions, so selecting it was a guaranteed error. Grok 4.3 remains available and working.
• Task-complete notifications now always appear. Windows notifications could silently fail to show when the AI’s summary contained a curly apostrophe. They now display reliably.
• Installer's Windows Defender exclusions now apply. The install step that adds Defender exclusions for smoother performance silently failed in previous versions; it now works correctly.

Performance

• Smoother on machines with third-party antivirus. Chat data is now written on a background thread, so real-time file scanning can never make the interface stutter — noticeably better on Bitdefender, Kaspersky, McAfee and corporate-protected machines.

v2.8.1 July 13, 2026

New

• Cortex Desktop source code is now on GitHub. Core components of the IDE are public on GitHub — read them, audit them, star the repo, open issues. Additional components will be opened as they clear security review, with full open-sourcing as the goal.
• Agentic Loop Mode is now a real switch. The Settings toggle now directly controls the autonomous verified-loop engine: goal → baseline → fix → verify → revise, repeating until real tests/lint/build pass and a second AI approves the diff. With the toggle off, the AI works one task at a time and cannot start a loop.
• Loop budgets are enforced for real. Autonomous loops now track their actual token and dollar spend per model and stop themselves the moment a budget is exceeded — alongside the existing iteration, no-progress, and time limits. A loop can never run away with your API bill.

Fixed

• Sign-in could fail after a retry. Clicking Sign In a second time while the first attempt was still waiting could make the login fail even after you approved it in the browser. Retried sign-ins now complete cleanly.
• Run File ran the wrong file. Pressing Ctrl+F5 could run the file selected in the Explorer instead of the tab actually open in the editor. Run File now always asks the editor which tab is really active.
• Signing in now unlocks MCP immediately. After logging in with a Pro account, the MCP Servers panel stayed locked until you pressed refresh. It now unlocks and connects your servers automatically the moment you sign in.
• MCP connections retry on flaky networks. A momentary network hiccup used to leave an MCP server disconnected for the whole session. Failed connections now retry automatically with increasing delays.
• The AI actually uses your connected MCP tools. The model could see a connected tool (like fetch) and still claim it "cannot access URLs." Its instructions now include a worked example, and it calls the tool instead of talking itself out of it.
• Usage & Billing now matches the website. OCR pages, embeddings, and web searches shown in Settings were device-local counts that drifted from your account. The desktop now shows the same account-wide numbers as cortex-ide.app, and a server-side metering bug that zeroed some token counts was fixed.
• Switching projects no longer mixes chats. If the AI was mid-task when you switched projects, its leftover output could overlap the new project's chat as garbled text. Switching now cleanly cancels the old project's in-flight work and fully clears the panel.
• Enabling Loop Mode no longer breaks editing. A wiring bug routed the Loop Mode toggle into an unrelated permission system that blocked the AI's file edits outright while the toggle was on. The toggle now controls only the loop engine; edit approval works exactly the same either way.

Performance

• Settings no longer freezes the app. Opening the Profile / Usage & Billing pages made three network calls on the interface thread, freezing the whole IDE until they finished. Settings now shows cached data instantly and refreshes quietly in the background.
• Faster streaming. Removed hidden logging overhead that wrote to disk three times for every streamed word of an AI response.
• Cleaner dark mode in Settings. Removed the last foggy gradients and glow washes — Settings is now the same pure flat dark as the editor and sidebar.

v2.8.0 July 12, 2026

New

• MCP Servers — connect external tools. Cortex now speaks the Model Context Protocol, the same open standard used by Claude Desktop and Cursor. Connect databases, GitHub, web search, browser automation, and hundreds of other tools, and the AI can use them directly inside your chat. Paste any standard mcpServers config in Settings → MCP Servers. Included with a Pro subscription.
• Live Preview. Open any local HTML file in a panel right beside your code, rendered by Cortex's own engine — no external browser needed. It reloads automatically every time you save, and the AI can render and test the pages it builds (checking the rendered content and the JavaScript console) as it works.
• Open Folder shortcut. The empty Explorer now has a clean "Open Folder" button with the Ctrl+O shortcut shown below it, so starting a project is one click away. The Explorer panel is also a bit wider by default so folder and file names aren't cut off.
• Refreshed brand. A new Cortex logo and mark throughout the app — icons, editor welcome screen, and Settings, which now shares the same dark theme as the rest of the IDE for a consistent look.
• License agreement at install. The installer now presents the Terms of Service, Privacy Policy, and EULA for acceptance before installing.

Performance

• Large chats open dramatically faster. Conversations that contained very large file diffs used to freeze the app for tens of seconds on open — a single big change that took around 40 seconds to display now renders almost instantly.
• No more startup freeze. Reopening a project with a long history no longer locks the window; the interface stays responsive while your chat fills in progressively.
• Reopen where you left off. Cortex now lands on your most recent messages the moment a project opens — scroll up to load older history on demand, instead of loading weeks of chat all at once.
• Smooth scrolling in long chats. Scrolling back through history no longer causes multi-second stalls.
• Calmer startup. Project indexing now waits for a quiet moment instead of competing with everything else while the IDE is opening.
• Settings opens instantly after the first time. Every time you opened Settings it rebuilt the whole page from scratch, showing a blank white flash before slowly filling in. It now stays ready in the background — the first open is unchanged, every one after that is instant.

Fixed

• MCP servers failing to connect after installing. Connected MCP servers worked in testing but failed to connect for everyone after installing Cortex normally. This is fixed — MCP servers now connect reliably in the installed app.
• Chat occasionally failing to respond. A server-side issue could cause repeated errors when sending a message. Cortex now backs off gracefully instead of hammering the server, and the underlying cause was fixed.
• Chat from the wrong project. Switching projects could show one project's conversation under another. Each project now keeps its own chat, cleanly separated.
• Interactive question cards. When the AI asks you to pick between options, it now always shows a proper clickable card instead of occasionally printing raw text.
• Hard-to-read text in Settings. Placeholder text in a few input fields, and the description text on the locked MCP Servers card for non-subscribers, were too faint to read comfortably. Both are now clear and legible.
• "View plans" opened inside Settings. Clicking to view subscription plans from the MCP Servers page took over the entire Settings window instead of opening your browser. It now opens in your default browser as expected.

v2.7.6 July 11, 2026

Fixed

• Update alerts now really appear. A display error could stop the "new version available" window from opening even after the update was detected. The window now opens reliably in both dark and light themes.

v2.7.5 July 11, 2026

Fixed

• Blank panels on some computers. On certain graphics hardware, the file explorer, code editor, terminal, and Settings could appear completely empty while the chat still worked. Rendering now uses a configuration compatible with all GPUs — every panel displays correctly on any Windows 10/11 machine.

v2.7.4 July 11, 2026

New

• Message timestamps. Every chat message now remembers exactly when it was sent — times are preserved across closing and reopening the IDE.
• Smarter project memory. MEMORY.md now stores distilled knowledge — completed tasks, files changed, AI-written session summaries — instead of copies of chat text, with one clean entry per session that updates in place.

Performance

• Faster streaming on long responses. Rendering cost no longer grows as a response streams — big answers now stay smooth from first word to last.
• Faster startup with large chats. Reopening a project restores your recent conversation immediately; older messages load on demand with "Load older messages".
• Lower memory usage. Internal pages share one render process and the semantic search index uses 8× less RAM — Cortex runs noticeably lighter, especially on 8 GB machines.
• Faster semantic search. Project-wide "find by meaning" queries respond much quicker on large codebases.
• Calmer background indexing. Project indexing no longer rewrites its entire index every few seconds — large projects index quietly without slowing down the rest of the IDE.

Fixed

• Saved API keys survive restarts. In installed builds, keys saved in Settings → Models & Providers could appear blank — or stop working — after closing and reopening the IDE. Key storage was rebuilt to read Windows Credential Manager reliably; keys now show masked and keep working after every restart.
• Settings page rendering as unstyled text. On slower machines the Settings window could open as a plain white/black text page with no controls. It now always loads fully styled.
• Startup screen behaving like a separate app. The loading mask covered the entire screen as its own window; it now lives inside the Cortex window and only covers Cortex.
• Ghost "capsule" window. A small floating pill-shaped window could appear over the IDE when sending the first prompt of a fresh session. Gone for good.
• Semantic memory index self-heals. A corrupted index (e.g. after a crash) no longer fails on every launch — it's rebuilt automatically, and saves are now crash-safe.
• Quieter, faster startup. Removed repeated key-decryption retries and per-message log writes that slowed launch and streaming on some machines.

v2.7.2 July 10, 2026

Fixed

• Update notifications now appear. Fixed a bug where the in-app "new version available" dialog silently never showed — including required updates. From this version onward, Cortex reliably notifies you when an update is ready and can install it in one click.

v2.7.1 July 10, 2026

New

• Anthropic (Claude) provider. Add your Anthropic API key in Settings → Models & Providers and use Claude models directly in Cortex.
• Thinking & reasoning display. Reasoning models (Anthropic and OpenRouter) now stream their thought process into a dedicated, collapsible thought card — separate from the final answer.

Fixed

• Saved API keys are now always used. Fixed a bug where keys saved in Settings → Models & Providers could be silently ignored after installation, making a provider appear "not working" even though the key was correct.
• Light mode readability. Tables, code blocks, and syntax highlighting now use proper dark text on light backgrounds — no more invisible white-on-white code in light theme.
• Theme switching both directions. Switching dark → light and light → dark now restyles every part of the chat correctly, including tables and code rendered before the switch.
• Project file search. Fixed a case where the AI agent's file search returned no results for patterns like **/*.py.

v2.7.0 July 9, 2026

Initial public release. Cortex AI IDE launches for Windows 10/11.

Highlights

• Agentic AI coding. An AI agent that reads your project, plans, edits files, and runs commands — with your approval at every step.
• Bring your own keys (BYOK). Connect MiMo, DeepSeek, OpenAI, Qwen, OpenRouter and more with your own API keys — stored encrypted in Windows Credential Manager, never on our servers.
• Full IDE. Multi-tab editor with syntax highlighting, file explorer, integrated terminal, and project-wide search.
• Project memory. Cortex remembers project context and your preferences across sessions.
• Light & dark themes, Markdown/table/diagram rendering in chat, and one-click code apply.