Agentic integrity · Open source · Apache 2.0
QuorX is an AI coding agent whose work is checked. It builds on your own computer with a real shell and a real browser, while decider models answer its questions, a separate reviewer scores every build, and a safety layer guards every command. Any model, from OpenRouter, the makers' own APIs, your subscriptions or your PC, priced before you pick.
Agentic integrity
What the agent does is tested by something other than the agent. And you can see how each model holds up before you trust it with a project.
When it's unsure, the agent writes down its own guess, then fast decider models vote. You see whether asking changed the answer.
The quorum ↓A separate model opens what was built, plays it and scores it before the reply. Problems go back to the agent.
The checks ↓Every command is parsed and classed before it runs. Risky ones wait for you, code QuorX can't read goes to the deciders, and system changes are refused.
Safety ↓A built-in build benchmark with hidden checks scores models on real tasks, and your own pass rate sits next to each model's price.
Models and prices ↓Across every model you can pick. Per-model results, the price before you pick, and model cards with each model's limits, quirks and QuorX results, so you can see which models build well and improve how you use them.
Watch the walkthrough · 1:36
The video loads from YouTube's no-cookie player only after you press play.
A game, a site, a desktop app, a fix. A prompt enhancer spells out what the request implies and shows it under your message, so you can correct it.
A real terminal, a real browser and the toolchains already on your computer.
Every file is checked as it is written. Then it opens what it built, plays it like a person would, and a separate reviewer model scores it.
Every change, commands included, with one-click Undo.
The quorum
When the agent is unsure, it writes down its own guess first, then asks a quorum of fast decider models yes/no or multiple-choice questions. You see whether asking changed anything.
When the snake reaches the edge of the board, what should happen in this game?
The agent writes down its own answer and how sure it is before the deciders vote, so you can see whether asking changed it.
When the agent would normally stop to ask you something, the deciders answer its question instead.
A standing set of questions goes to the deciders about every build too, not only when the agent is unsure.
Safe by default
Every command line is parsed and each part is classed. Your rules decide what runs; the rest comes to your approval card. Starting a server gets the same check.
Strict is the default and asks you about anything that isn't read-only. YOLO runs most commands on its own, after a snapshot; pushes, deploys and writes outside the project pause for you, and commands QuorX can't read go to the deciders.
The project is snapshotted before each command, so Undo covers what a build script or an rm did, not just the agent's edits, even when commands ran at the same time.
A command that deletes many files outside the project is stopped. A cap per reply, an optional daily cap, project budgets and a media cap. Stopped and failed calls count too, so caps match what OpenRouter bills.
Example values from one chat's settings. You set your own.
Each command runs in a Windows job, so Stop ends every process it started and confirms 0 processes left. Pressed before a command starts, Stop means it never runs.
System changes (services, the registry, disks, admin rights), deleting your profile, a drive root or anything above the project. No rule or mode can approve these.
rm -rf ~your profileRemove-Item -Recurse C:\a drive rootrm -rf ../..above the projectStop-Service WinDefenda system change
Hundreds of models through one OpenRouter key, the makers' own APIs with your key, the plans you already have, or Ollama and the AI apps on your PC. The picker starts with where a model runs, then searches, sorts by price, filters by what a model can do, and shows your own pass rate next to each price.
Every step and every cent is on the record.
$0.01095for one reply, 16 model calls, itemised.
One key, every makerpay per call
The makers' own APIsyour key, their prices
Subscriptionsa key with plan limits
Apps on this PCyour sign-ins there, no key here
Same tools, approvals, Undo and Stop on every one. Mix them: one model for the chat, another for the reviewer, a third for the helpers.
Each one finds its toolchain, compiles what a reply changed, runs and drives the app, and debugs with breakpoints.
Your terminals, what's running and your servers in one place, with panes you can stack and resize, and Stop for everything.
Each chat works in its own project with its own budget. When a budget runs out, the reply stops and asks.
Point at part of a page to say what should change. Export as PDF, PNG or a ZIP.
A real browser tab, or your own Chrome, Brave or Edge through the QuorX extension; the desktop apps you allow, at View, Click or Full; Android emulators and interactive terminal programs.
What's new in each version, plus reports and research articles. New pieces arrive between versions.
The agent uses any MCP server's tools, and any MCP client can hand work to QuorX. Write or import your own skills, for every project or just one.
What's new, release by release
0.2.19 is out now as the installer and from source; the Microsoft Store copy updates itself with a later Store release. Pick a release to see what it added, with a guide for each big feature. All articles →
QuorX reads each model's limits and builds the request to fit, instead of failing and retrying: reasoning that can't be turned off, the levels it accepts, the tool settings each provider takes. On OpenRouter, 507 of 2,029 tool endpoints refuse "required".
Guide →A tool card appears while the call is still being written, naming the file and how much is done. It works with every provider, from OpenRouter and the makers' APIs to Ollama and the agents on your PC.
A Continue button carries on from where you stopped it, and "go on" continues instead of starting over.
Small slips in a tool call are repaired, and ones that don't fit go back with each wrong field named.
QuorX waits as long as the provider asks, and the reply says how long instead of looking stuck.
A conversation too long for the model is trimmed and tried again. A huge tool result keeps its start and end, with the rest saved for the agent to read.
GuideWhat's new in 0.2.18: smooth with many heavy chats
ResearchThe chats that never left: how the slowdown was measured and fixed
A leak kept every chat you had ever opened in memory. After 6 long chats, 666,710 page parts were alive with about 5,000 on screen. Now it's 15,252, and the window went from 511 MB to 338 MB.
The story →A tab you can't see sleeps after 5 minutes and wakes at the same scroll in about a second. Each one cost about 127 MB. Tabs with sound, a half-filled form, the agent at work or DevTools stay awake.
Guide →A 208-message chat froze the window for 1.8 s. Now it opens in 130 to 340 ms: the newest messages first, older ones as you scroll up.
Minimised with nothing running, QuorX uses 0% of the processor.
Memory, browser tabs, processor and stutters in a small box, with Copy for a bug report. Off unless you turn it on.
Pictures, clips, sounds and music a tool makes show under their own step as they land. Also: empty the agent's folder to the Recycle Bin, and model chips show where a model runs.
The desktop app runs on Electron 44.10.2, which brings the October 2026 Node security fix.
Renders, ffmpeg steps and project scripts are checked for harm only, so a video's last steps no longer get skipped. A command you approved runs again without another vote.
Sidebar chats show their state in their own colours: working, waiting for you, using Chrome or the desktop, finished or stopped.
The agent can pick a voice and speed, and a size, quality, seed or sound for pictures, clips and music.
244 animated pixel art effects, and canvas games fill any screen with a full-screen button.
A new provider under Makers' APIs, with its balance in Settings.
Top to bottom, in the order things happened, with a time on every message. No bubbles: full-width rows with a glowing edge, yours teal, the agent's violet, turning green, amber or red with its state. A finished reply folds its work into How it got here.
Guide →It reads a file again if you changed it, hears when background jobs end, reads other tools' notes, loads rare tools only when needed, and runs your hooks around any tool. Helper roles and typed memory, all in Settings.
Guide →Allow one ffmpeg command and every ffmpeg command in that chat runs without a card. Always allow is on every command card.
Guide →The reply, media and daily caps and project budgets become one budget per chat. /goal is the one way to ask for a plan.
Files, Review changes, the live plan and long answers open in the side pane's Artifact tab; Compare splits it.
Two built-in skills for films with exact easing and timing, including HyperFrames. Videos in your projects can be scrubbed.
GuideMCP servers: give the agent more tools
GuideYour own API keys, and one Providers page
GuidePermissions, in one place
GuideChange QuorX's settings from a chat
GuideA calmer, faster QuorX
GuideSub value: what your subscription covers
The agent uses any MCP server's tools, a program on your PC or a server on the web. Add one in Settings, paste its setup into a chat, or bring in the ones your other tools already have. Notion, Linear and Sentry sign in from your own browser.
Guide →OpenRouter; the makers' own APIs with your key (OpenAI, Anthropic, Google Gemini, Moonshot Kimi, Meta); OpenCode Go and Command Code plans; and Ollama, OpenCode, Claude Code and Grok Build on your PC. Each with its own page, spend and a free test.
Guide →Every rule, app, folder, site and tool you allowed, by chat, project or everywhere, with how often each ran. Switch any off, or remove it.
Guide →They say what will run and why it asks, and "Allow always" asks what it covers and where. Type instead of clicking to answer in words.
Guide →Ask the agent to read or change QuorX's and the project's settings. Every change waits for your click, and keys stay out of reach.
Guide →What your subscription calls would have cost at list price, per call, reply, chat and in total, so you can see what your plan is worth.
Guide →Every server, command, terminal, browser tab, Chrome and desktop session QuorX has going, in every project, each with Stop.
Tour →Two or three tabs at once, stacked and resizable: the Browser over the Terminal, Steps over Files. Remembered per chat.
Tour →Pictures, videos, PDFs and Markdown show in the chat. The message box gives your text the full width, with one row of controls.
Tour →QuorX shrinks beside the app it's using, with a Stop banner on top. Computer use and Chrome start only after your click.
Guide →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 in your own browser.
GuideYour own AI apps as the chat model
GuideReplies that finish the job
GuideHow to use computer use
Run a chat's main model through OpenCode, Claude Code or Grok Build on your PC. Your own plan pays, QuorX holds no key, and its files, browser, terminal, checks and approvals work the same.
Guide →Chats, the helper, the reviewer, the prompt enhancer and fresh eyes can all run on an app on this PC. What needs OpenRouter is switched off, and says so.
Guide →The model picker starts with the source, each with its logo, then that source's own filters, so a long list stays short.
A reply that says what it will do but makes no call is sent back to make it. One that loops or runs far too long is stopped, keeping its calls.
Guide →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 a file can't make one.
Guide →One held stroke through many points, smooth held moves, and up to 40 steps in one go.
Guide →Next to 👁, 🧭 follows the running reply's steps in the side pane, for that chat only.
Rename, Duplicate and Delete from one menu on each chat row, also on right-click.
GuideSchedules that run themselves
GuideEach chat's own notes, and folders outside the project
GuideBlock a provider that breaks your model
A chat that runs every weekday at 9, every hour, or once tomorrow. Press ⏰ Repeat this under a good reply, pick when it tells you, and three failed runs in a row pause it.
Guide →Allow each app at View, Click or Full, and the agent works in apps already open or opens installed ones by name. A glowing edge and a Stop button show while it works.
Guide →A countdown to each next run, a run going on now glows in the list and pulses in the sidebar, and each schedule has one chat.
Guide →Stop OpenRouter sending calls to a provider that breaks tool calls, for every model or only some (deepseek/*).
Every chat keeps its own agent notes, and a chat can stay off its project's notes until you bring it in.
Guide →One card adds a folder to the chat, and the file tools work there. Paths in replies open in File Explorer.
Guide →Always allowed rules turn on and off and count their uses, in the chat, project settings and Settings.
Rules match exactly what they name, and dialogs such as "Replace it?" ask you first.
A film starts with a brief, a story pick and a storyboard, then music first and moving pictures, before anything is spent.
Each category laid out for what it controls. Storage shows what QuorX takes and clears it, Skills you write work in every project or just one, and News arrives between versions.
QuorX opens at quorx.localhost, so your sign-in stays away from the apps you build and run on localhost.
Big projects, long replies and large backups no longer slow QuorX down. A streaming reply keeps its text in place, so you can select it as it arrives.
Starting a server is checked like any command, reading your keys asks first, and the agent's keys and clicks go only to apps it started.
The agent's tabs stay open and the next reply carries on in them. ⚡ YOLO can act in your own tabs without asking, if you turn it on. Pairing shows a code to match.
Good to know when you update from 0.2.11 or earlier
.gitignore ignores and build folders.The parts that never change come first, so more of each step is read from the provider's cache. On our benchmark run it passed 2 of 2 tasks vs 1 of 2, at $0.094 vs $0.127 a run. That was one run of each version, so most of the gap is run-to-run variation.
How it was measured ↗Tougher build tasks with hidden checks the agent can't see, and stricter scoring for originality.
Recommended models get a card with their limits, quirks, provider tips and QuorX results, and the agent reads the ones it uses.
Usage shows how much of each model's input came from the prompt cache, so you can see where the money goes.
/review lists bugs in what changed with file and line. Hooks run your own commands around the agent's work. Schedules start chats on their own. Memory keeps facts you approve.
A short tour of projects, Strict and YOLO, the deciders and costs. Export saves your chats, projects and settings to one file; Import restores them anywhere.
Built with QuorX
Two apps from real QuorX 0.2.12 runs, each from one request: written, tested and reviewed by QuorX, and hosted here as they came out.
Live weather and air quality for any city, a sky drawn from the real conditions, and Kite Day: fly a kite in the city's real wind. QuorX found the free weather API by researching it in its own Chrome tabs, then built the app.
Uses live data from Open-Meteo once you open it.
A black hole you can play with: gravitational lensing ray-marched on your graphics card, a Doppler-beamed disk, and probes on relativistic orbits that precess into rosettes.
Heavy on graphics: it checks your GPU before it starts.
Deliberately narrow
Research · Video direction
Ten video skills, one per route: code, images, AI clips, footage, or a mix.
Analysis only Nothing here has been tested on video yet.
Newsletter
One email per release or research piece. Kept on our own server, never sold, and you can leave anytime.
Which one updates first?
Windows
Microsoft signs it and keeps it up to date, so Windows shows no warning.
0.2.19 isn't in the Store: get it with the installer below or from source. The Store copy updates itself with a later Store release.
Direct download (.exe) QuorX-Setup-0.2.19.exe from the 0.2.19 release. It updates itself, but it isn't code-signed yet, so Windows SmartScreen may warn the first time you run it. Every release has SHA256SUMS.txt and a signed build attestation:
PS> Get-FileHash .\QuorX-Setup-0.2.19.exe -Algorithm SHA256
Open source · Apache 2.0
QuorX is built in the open. Run it from source, tell us what breaks, suggest what's next, or send a pull request. Every bit helps.
Run it from source on Windows, with Node.js 20 or newer and an OpenRouter account:
git clone https://github.com/mims1234/QuorX.git
cd QuorX
npm install
npm startThen open http://localhost:5188 and choose Sign in with OpenRouter.
Prefer the desktop app? From the same folder (the first line installs Electron, about 100 MB, once):
npm run desktop:setup
npm run desktopOfficial downloads come only from the Microsoft Store, GitHub releases and this site. Forks are welcome under a name of their own.