What's new

The agent can use the tools of any MCP server, on this PC or on the web: add one in Settings, or paste its setup into a chat.

Web MCP servers that need an account (Notion, Linear, Sentry and others) sign in from QuorX, in your own browser.

MCP servers already set up in your other tools are found and offered, so adding them is one click.

Direct API keys for OpenAI, Anthropic, Google Gemini, Moonshot Kimi and Meta, with their models, prices and costs.

Settings has one Providers page: every provider with its own page, and the defaults for new chats at the top.

Command Code and Ollama on this PC can run a chat's model, and OpenCode Go works with your own API key.

Sub value shows what a call on a subscription would have cost at list price, so you can see what your plan covers.

Ask the agent to read or change QuorX's and the project's settings; every change waits for your click.

One Permissions window for every rule, app, site and tool you allowed, by chat, project or everywhere, easy to remove.

Approval cards say plainly what will run, what "Allow always" covers, and where to take it back.

Running in the sidebar lists every server, browser tab, Chrome and desktop use QuorX has going, in every project.

The side pane shows two or three of its tabs at once, stacked and resizable, like a split terminal.

Project settings has a short main window, with Design system, Hooks and Permissions each in its own window.

Right-click a project in the sidebar for its quick actions: new chat, rename, pin, open the folder, archive, delete.

A compact message box: the text gets the full width, with one row of controls that folds as the window narrows.

Pictures, videos, PDFs and Markdown files show inline in the chat.

The Browser tab has volume and mute for every tab, the agent's alone, or one tab.

A click on a link opens it in the Browser tab; Ctrl+click opens your own browser.

Computer use works on one screen, with QuorX compact beside the app it's using and a Stop banner on top.

Computer use and Chrome only start after you click to turn them on, and turning Chrome off mid-reply stops it.

Typing while an approval card waits sends your words to the agent, instead of denying.

Steps stays fast on replies with thousands of steps, and opened sections stay open as new steps arrive.

Notifications are short, and you can pick All, Needs me or Off right in the What QuorX may use panel.

Save keeps Settings open, and the theme is picked in Settings instead of a sidebar button.

Project costs say why Spent can be more than the chats add up to: deleted or rewound chats, or replies stopped early.

Calls in DeepSeek's own format, or copied from earlier ones without the step's key, are asked for again instead of ending the reply.

File writes with quotes in their text (CSS fonts) and calls with flat arguments now run, instead of failing again and again.

A long build no longer ends on a third blank step: a blank answer gets its retries again after any step that made calls.

Menus and popovers over the Browser tab are never hidden behind the page.

The agent's page can't stay stuck on top of other pane tabs, and tabs left playing after a reload are closed.

Sending to an @ chat shows your message at once, sends it once, and the model knows what "this" means.

MCP servers stop fully on Windows, one that floods its output is stopped, and web sign-ins renew reliably.

MCP setup files, which can hold keys, are off limits to the agent's file tools, like your other key files.

A new Quorium saved without an id is kept, and Ollama on another host or port is found.

The Permissions window opens fast with many chats, and bad requests get a clear error instead of a server error.

Run a chat's main model through OpenCode on this PC, with the providers you connected there: your plans pay, and QuorX keeps its own tools.

OpenCode's own free models answer only in OpenCode's app, so they show in the picker but can't be picked.

Helpers, the reviewer, the prompt enhancer and fresh eyes can run on a model from an app on this PC too, even with no OpenRouter key.

A model your account settings there don't allow (OpenCode Go's Global regions) says so at once, instead of a server error after retries.

Claude Code on this PC: a chat's main model through the Claude Code you are signed in to.

Grok Build on this PC: a chat's main model through Grok, for text work (🖥 Computer use is off with it, and the chat says why).

Chats run without an OpenRouter key on a model from an app on this PC; what needs OpenRouter is switched off and says so.

The model picker starts with where a model runs (OpenRouter or an app on this PC), each with its own logo, then that source's filters.

Chat rows have one ⋯ menu, also on right-click: Rename, Duplicate and Delete.

🧭 next to 👁 follows the running reply's steps in the side pane, for that chat only.

Computer use draws freehand: one held stroke through many points, and up to 40 steps in one go.

A reply that says what it will do but makes no call is sent back to make it, and a model that keeps stopping short ends with a clear note.

Calls a model writes inside a code block or right after a sentence are asked for again instead of showing as raw text.

A reply that loops, writes the conversation's next turns itself, or runs far too long is stopped, keeping the calls it made before.

A failed call through an app on this PC stops the reply and says which model failed and why; falling back to OpenRouter is your choice in Settings.

Tool calls written as text run only with this step's key, on a line of their own, outside code, quotes and plans, so a page or file can't make one.

A plan or a mention of a call is never asked for as a call, and Claude Code's, Grok's, OpenCode's and Cline's sign-in files stay out of the agent's reach.

The Stop banner's quotes and other non-English characters no longer come out garbled during computer use.

A final answer that reads like a next step is kept as the answer, without a false warning.

Computer use: allow apps per chat at View, Click or Full, and the agent works in apps that are already open or opens installed ones by name.

A glowing edge and a Stop button while the agent uses your desktop or one of your Chrome tabs.

Schedules as jobs: manage them from a chat, Repeat this on any reply, notify rules, auto-pause after three failures, and run dots.

Each schedule has one chat, schedules have a project of their own, and they can repeat as often as every 3 minutes.

Schedules show live: a countdown to each next run, and a run going on now glows in the list and pulses in the sidebar.

Blocked providers in Settings, for every model or only some (deepseek/*); InferenceNet is blocked for DeepSeek models by default.

Always allowed rules turn on and off and count their uses, in one view in the chat, project settings and Settings.

Each chat has its own agent notes, and a chat can be kept off its project.

Schedule chats are marked, and a scheduled turn says it sees only its prompt.

Open chats pick up new scheduled runs and their cards; a schedule’s card waits 15 s (settable), and an unanswered one ends the run.

A scheduled run’s Chrome tabs close 30 s after it goes well; a failed one keeps them for you to look at.

A scheduled run cut short counts as failed, with no out-of-steps card.

Folders outside the project: one card adds the folder to the chat, and the file tools work there.

Paths to local files in replies open in File Explorer.

Desktop actions are quicker and steadier: windows settle before the next step, typing reads itself back, and long typing no longer times out.

A reply keeps every step it took, and replies saved short get them back.

Approvals do what they say: rules match what they name, scratch scripts and cmd /c are checked, and waits are safer.

Only system tools and admin shells are never used on the desktop; what you named in your message is your go-ahead.

A decider’s answer can’t claim you did something, and Replace it? dialogs ask you first.

File Explorer is found by its Start-menu name, with its open folder windows.

White screenshots, scaled displays and apps that outlive QuorX are handled.

Turning off the cost badge hides the Spending total too.

Saves retry a file Windows has locked, and the browser recovers when its context is gone.

Video Mode: a film starts with a brief, a story pick and a storyboard, then music first and moving pictures, before anything is spent.

Settings, redesigned: each category laid out for what it controls, with Reset category to bring it back to its defaults.

Storage in Settings: see what QuorX takes on this computer, clear it, and compact older chats.

Skills: write, import, edit and trash your own skills, for every project or just one.

News from the website: new articles, reports and What's new arrive between versions.

A slim line under every message box shows the mic key, what it does, and Hold or Toggle.

⚡ YOLO can act in your own Chrome tabs without asking: a setting, off by default, with a note for every step it takes.

Up to 500 steps a reply; at the limit, a card asks how to carry on.

A streaming reply keeps its finished text on the page, so you can select it and code blocks stay put.

Big projects and long replies no longer freeze QuorX: snapshots, search and backups now work in the background.

Search skips what .gitignore ignores and build output, and says when it stopped early.

Backups over 4 GB work, and very long replies use a tenth of the memory they did.

Chats you aren't looking at let go of their memory, and come back complete when you open them.

The agent's Chrome tabs stay open, the next reply carries on in them, and a doubtful step asks you instead of being refused.

The browser agent is steadier, can be stopped mid-browsing, and works with Chrome extension 0.3.

The same chat open in two windows: saves are merged instead of one window overwriting the other.

Pane tabs are icons only, with each tab's name and what it does in its hover tip.

Durations read as people say them: 21m 42s instead of 1301.5s.

An update button next to ✕ in Settings, and a live update strip in the sidebar.

QuorX opens at its own address, quorx.localhost, so your sign-in no longer reaches dev servers you run on localhost.

Starting a server gets the same check as any command, and Stop pressed before a command starts means it never runs.

Reading your keys by any route asks you first, and private keys are hidden from the agent in full.

Keys and clicks from the agent go only to apps it started, never to your own windows.

Exporting your API keys asks once more, naming the folder they go to.

Pairing the extension shows a code the extension shows too, and the key stays with that extension.

Attached documents reach the agent again.

The project list survives a crash mid-save, and a file in use no longer stops QuorX.

Stopped and failed calls now count toward spend, so budgets match what OpenRouter bills.

Undo covers commands that ran at the same time, and names any file it can't cover.

The terminal no longer repeats itself after a reconnect, and Enter while typing Chinese or Japanese doesn't send.

Copy in the code view gives the code exactly as written.

Smaller fixes: News sync errors, unreadable folders, memory saves on a busy disk, and the desktop Browser tab's placement.

The QuorX extension for Chrome, Brave or Edge: tick 🧩 Chrome in a chat and the agent works in your own browser, signed in as you.

Give one of your browser tabs to a chat by dragging it there; the agent works only in that tab, and Take back ends it.

The Terminal tab holds your terminals, what's running and your servers, with terminal panes you can stack and resize.

✨ Enhance rewrites what you typed clearer and fuller before you send it. Press it again to go further, or ↶ to undo.

Drag files, browser tabs or terminal output onto a chat, or right-click them → Add to chat.

Type words in the Browser tab's address bar to search, with Google, DuckDuckGo, Bing, Brave, Startpage or your own.

The agent can list, rename and remove the servers saved for a project, not only add them.

⚡ YOLO runs most commands on its own; only extreme ones pause for you, in warning colours.

Safety & commands settings show the simple choices first, with every detail under Advanced.

Replies read in order: the plan sits above its Go ahead, each review where it happened, answered questions on one line.

Your terminal is a normal terminal, and tools you just installed are found at once.

Follow also switches to the Terminal tab when the agent runs a command.

Links in chats and News open with Ctrl+click, in the Browser tab.

Closing a tab that shows a running server asks whether to stop the server or keep it running.

The agent knows the names of your project's secrets, never their values.

Guarded against the October 2026 Node.js security issue until the desktop app's fix ships.

Screenshots of scrolled pages are no longer blank, so the agent stops re-checking pages that are fine.

The agent's screenshots no longer stall while its tab is hidden in the desktop app.

Fresh eyes no longer raises an alarm about a file the agent hasn't written yet.

After OpenRouter sign-in, the browser tab says to close it and QuorX comes to the front.

Smaller fixes: the sidebar's title row, the ▶ in Servers, Clear output next to Close, and the Running stop button.

QuorX is in the Microsoft Store: Microsoft signs that copy and keeps it up to date, so Windows shows no warning.

The agent's browser tells it whether a page started any sound, so it can check a game or a player it can't hear.

The Windows app is about 55 MB smaller: it carries only the terminal parts it uses.

The Models report in News points to OpenRouter's rankings for choosing a model.

With another copy of QuorX already running, the desktop app starts its own instead of opening the other one.

News has seven new articles on what 0.2.7 measured and learned, from the cache-friendly prompt to a lock-down post-mortem.

The Models, Tests and Platforms reports in News are rebuilt with 0.2.7's benchmark and test results.

The agent's browser can wait for a timer, a round or an animation, instead of sleeping in the shell.

The agent's browser presses key combinations like Shift+R and Control+S, and a capital letter holds Shift.

The Tests report in News shows each pack's results only, not details of the computer that ran the tests.

/review lists bugs in what changed (the last reply, uncommitted changes or the branch) with file:line; send the ones you agree with to the agent.

Hooks run your own commands around the agent's work, like npm run lint after every file write, set in each project's settings.

Schedules start a chat on their own, like every weekday at 9 in a project, with the prompt, model and mode you saved.

Memory: a short list of facts and preferences you approve, like "I use pnpm", that every chat gets; edit it in Settings.

Export my QuorX saves your chats, projects, settings and rules to one file, and Import restores it on any computer.

Paid media work with several parts is priced up front as one production plan and approved once.

A video clip can start from a picture, and edit_video joins clips, stills, voice and music into one .mp4 with ffmpeg.

Recommended models have a model card with their limits, quirks and tips, and the agent reads the ones it uses.

A short first-run tour shows projects, Strict and YOLO, the deciders, costs and the terminal.

Research notes for 0.2.7 in docs/research: how QuorX checks its agent across models, a cheaper prompt, and a 0.2.5 post-mortem.

Each step's prompt is smaller and puts what doesn't change first, so more of it is read from the cache.

Usage shows how much of each model's input came from the prompt cache.

Closing the desktop window explains that QuorX keeps running in the tray, and Quit QuorX is easy to find.

The desktop app ships its full terminal again, plus the terminal view and math files the installer had left out.

A reply resumed after a restart stays in one bubble and keeps its result.

Two copies of QuorX on one computer no longer sign each other out.

A project inside another project's folder no longer fills that project's file list and prompts.

A file check that runs out of time no longer reports a syntax error in a file that is fine.

QuorX answers only the computer it runs on: phone and other-device access is gone, and so is everything it opened up.

Hook commands go through your command rules and approvals like the agent's own, and /review never sends secrets files.

An imported backup never changes your approval settings, and its schedules arrive paused.

Every pull request is tested on Windows and Linux, and new versions reach installed apps only once promoted.

0.2.5 locked QuorX out of its own data files, so projects, chats and the Terminal seemed gone; nothing was lost.

QuorX repairs those files by itself when it starts, before it reads anything, with no administrator rights needed.

The desktop window no longer says its app link is not valid after a restart.

The data folder stays private to your account: only the folder is locked, and its files inherit that.

If the project list can't be read, QuorX says so instead of showing an empty list.

QuorX answers only its own windows: open the link it prints at start once per browser (the desktop app does this for you).

Other devices are off for new installs, and QuorX listens only on this computer until you turn them on.

Commands can't read QuorX's keys or browser profiles, and reading files outside the project asks first.

Web fetches check every redirect, and a page from one chat can't read another chat's files.

Releases are built by a read-only job and published by a separate step, with checksums checked in between.

Command rules move from each project's folder into QuorX's own data folder; add any you saved before again.

A video direction skill for trailers, explainers and social cuts, with the research behind it.

Code and styles for features that are gone were removed, about 1,100 lines lighter.

QuorX is open source under the Apache License 2.0: use it, change it and fork it, for personal or commercial work.

Each release lists the installer's SHA-256 checksum and a signed proof that GitHub built it from this repository.

A shorter README with the trailer, plus pages on getting started, features, privacy and security.

A project's spend in the sidebar no longer goes below zero while a reply is running.

The agent's own page in the Browser tab shows its picture after the reply, instead of a broken image.

Prompt enhancer: a small model spells out what your request implies, shown under your message for you to correct.

The agent and the deciders read that reading, so a movie trailer's voice-over counts as asked for.

In YOLO, a question or an unsettled approval waits 15 seconds for you before the deciders answer or it carries on.

Wait for me, or typing in the card, stops that clock; set the time in Settings under Budget & approvals.

A helper script the agent writes in its own scratch folder is judged on harm only, so the deciders no longer turn it down for not being asked for.

PowerShell lines that only work with text, such as $_.FullName.Replace(...), .IndexOf(...) and .Substring(...), count as reads and no longer wait for approval.

A plan that nobody has started after 10 steps gets a nudge to begin.

An approval the deciders couldn't settle showed up as two identical cards.

The placeholder left when old text is trimmed from the conversation could be written into a file. Such writes are refused now.

Bugs the reviewer finds that a user would notice send the build back even when its scores pass.

After the checks send a build back, the final reply keeps its answer to your message and adds what was fixed.

The cost card at the end of the steps is more compact.

The reports, articles and What's new use neutral wording, written for anyone who reads them.

Numbers worked out from the computer's own logs are labelled On this computer, and still never leave it.

QuorX has its own logo, a Q whose tail is an X, in the top bar, on Home, in Settings, by the version and as the tab icon.

The desktop app uses the logo for its window, taskbar, tray, notifications and installer.

News in the sidebar: what's new in each version, reports and articles, with a dot when something is new.

Design systems: pick one of six starters or measure one from a site or screenshot, and every chat in the project builds to it.

Comment on part of a page in the Browser tab: point at an element, say what should change, and the agent changes only that.

Export a page as a PDF or a full-length PNG, or the project's files as a ZIP.

Any MCP client can hand work to QuorX: list projects, start a reply, follow it and stop it; approvals stay with the person at QuorX.

Replay a finished reply step by step with its real timing, a scrubber and speeds up to 16×.

Home's start box picks a project, a recent folder or a template before typing.

Projects show their GitHub repo, can be pinned or archived, and the chat header shows what the project has spent.

Costs update live while a reply runs, and a reply that ended early still counts what its calls spent.

A simpler message box: the model, then This chat's settings with the caps, reasoning and every tool as one icon.

Home lists the projects, pinned ones first, each with its latest chats.

When the reviewer can't open any page, it says so instead of scoring a browser that failed.

The side panel can be hidden for a chat-only view, text size and page zoom are set separately, and notifications carry QuorX's name.

The agent works through a real shell, so it can build command-line programs and desktop apps, not only web pages.

Rules read every command before it runs; deleting the user profile, a drive or a system folder is blocked however it is written.

Stop ends everything a command started and confirms nothing is left, and Undo now covers what commands change.

A tripwire stops a command part-way when many files outside the project disappear.

In YOLO the deciders vote on risky commands; they see only the facts, need a quorum, and never include the chat's own model.

Spending caps per reply and per day, and installs for the whole computer always ask first.

QuorX checks that a toolchain can really build here before the agent starts, and says what is missing.

The agent can answer prompts in terminal programs, click through Windows desktop apps and Android emulators, and use a debugger.

A Running tab shows everything running in a project and who started it, with Stop on each and Stop all.

Long builds and installs move to the background instead of being cut off at a time limit.

Platform packs: one file per kind of app tells QuorX how to detect, check, run, debug and explain it.

Anyone can test the packs their computer supports with npm run test:packs, without an OpenRouter key.

.NET: WinForms and WPF apps are built and driven, dotnet test failures are summed up, and netcoredbg debugs them.

Scripts that run other programs are read first and always ask, so a wrapper script can't hide what really runs.

Tauri apps are built, and their window is read and clicked through WebView2.

Go, C, C++ and Rust programs can be debugged with breakpoints, and C and C++ projects get a build check before the reply.

Electron apps are driven through their DevTools port, and the debugger attaches to the main process.

Godot and Unity projects are checked and their tests summed up; MSIX and WiX installers are built with test certificates only.

Android emulators are started and driven, crashes point to the first line of the project's own code, and Flutter apps are checked and debugged.

Java and Kotlin: Gradle and Maven compile checks, JUnit summaries, and a debugger built on the JDK's own.

QuorX runs safely on macOS and Linux: Stop ends whole process groups, rules read bash and zsh, and key folders are off limits.

On a Mac, iPhone and Mac apps are built with Xcode, driven in the Simulator and debugged; crash reports are summed up.

Python desktop apps use the project's own virtual environment, and tkinter and PyQt windows are driven and debugged.

The update button in the sidebar is short and fills up like a progress bar while the update downloads.

With the sidebar folded, a waiting update shows as an icon above restart, theme and settings.

A reply no longer fails its final check when the hidden browser won't start; it uses the agent's own tab or checks the files.

QuorX finds Chrome installed for one user, 32-bit Chrome, Brave and Chromium, and says why a browser didn't start.

Pages the agent makes load from their own local address, so storage works in artifacts and project pages.

The side pane's tabs show only icons when the pane is narrow, and the servers button stays in view.

When the same problem comes back, a second model with a clean context finds the likely cause before the next try.

QuorX runs any model through OpenRouter, with a quorum of decider models behind its choices.

The agent builds web apps, sites and games, and opens each one in a browser to check it before it replies.