Cortex AI IDE was built by one person, in twelve months, for under $2,000 — and most of that money went to server-side testing, not development. Eight months of planning, four months of building at 14 to 16 hours a day, no team, no funding, no company behind it. It shipped its first public release — v2.7.0 for Windows 10 and 11 — on July 9, 2026, then shipped twelve more releases in the next ten days. This is the honest version of that story, including the part about why it exists at all: I don't think your prompts and your code should ever touch someone else's server.
Why build another AI IDE at all?
Because the tools I wanted to use kept failing the same tests. They treated Windows as an afterthought — Electron wrappers and "just use WSL." They asked me to trust a subscription with unbounded inference costs. And above all, they asked me to route my code and prompts through their servers and take their word for what happened there. The 2026 consolidation wave made the trust question sharper, not softer: the major AI IDEs are now business units inside giants, and your relationship is with whoever owns them this quarter.
There's an old model for the alternative, and it's the one that inspired me: when Linus Torvalds built Linux, he wasn't inventing computing from nothing — he took what existed, organized it, and gave it to the public on the public's terms. Cortex is my attempt at the same move for AI coding: the capabilities the big clubbed-together tools have — agentic loops, multi-provider models, MCP, semantic search — reassembled into one native Windows IDE that belongs to its users. Free to use with your own keys. No subscription required to code.
The architecture is the privacy policy
Cortex's privacy story isn't a promise in a policy document — it's how the software is wired:
- Cortex is the wrapper; the model is the brain. When you chat, your prompt and code go directly from your machine to the LLM provider you chose — Anthropic, DeepSeek, OpenAI, MiMo, Qwen, OpenRouter — over TLS. There is no Cortex proxy in the middle. Architecturally, my server cannot read your prompts or code, which is a much stronger guarantee than "we don't."
- Downloading is anonymous. Grab the installer and go — no account, no sign-up wall.
- Signing in reveals exactly one thing: your email. That plus your plan is all the server knows. What you're building, what you're asking, which files you open — nothing is tracked, because I never built the tracking.
- Your API keys never leave your machine — Windows Credential Manager, DPAPI-encrypted, with an AES-256-GCM local backup.
Why "trust me" isn't good enough — the Grok Build lesson
If this sounds paranoid, July 2026 provided the counterexample. A security researcher intercepted the traffic of xAI's Grok Build CLI and found it uploading entire Git repositories — full history, secrets included — to a cloud storage bucket, far beyond what any coding task needed: on a 12 GB test repo, ~192 KB of model traffic was accompanied by ~5.1 GB of repository upload. A canary credential planted in a .env file appeared verbatim in the captured traffic. The "Improve the model" toggle developers assumed was a data control didn't stop the uploads. After public exposure, xAI silently flipped a server-side flag — disable_codebase_upload: true — and Elon Musk pledged deletion of historical data; as of this writing there's still no formal incident report or way for developers to verify the purge.
The lesson isn't "xAI is uniquely bad." The lesson is that when your tool routes data through a vendor, you're trusting configuration you can't see, changeable remotely, by people you'll never meet. The only architecture that survives that lesson is the one where the data never leaves in the first place. That's Cortex's design — and it's why the code is being opened for inspection: core components are public on GitHub today, with full open-sourcing planned as each part clears security review, so third parties can verify the no-tracking claim instead of taking mine on faith. Fitting, for a product whose whole thesis is verify, don't trust.
What ten days of shipping actually looks like
Four months of development ends the day real users arrive; then the real education starts. The changelog is public and unvarnished:
- July 9 — v2.7.0. Launch: agentic editing, BYOK for five providers, full editor, terminal, project memory.
- July 10–11 — v2.7.1 through v2.7.6. Reality arrives: saved API keys silently ignored after install, panels rendering blank on certain GPUs, a ghost "capsule" window. Five releases in three days, fixing what testing on one machine can't catch — other people's machines.
- July 12 — v2.8.0. MCP support, Live Preview, and performance work that took a 40-second chat-open freeze to near-instant.
- July 13 — v2.8.1. Core source code lands on GitHub, and the loop engine's token/dollar budgets become real enforcement, not a checkbox.
- July 14–18 — v2.8.2 through v2.8.6. Antivirus stutter fixed with background-thread writes, vision-model image handling rebuilt, Claude Sonnet 5 and Kimi K3 added within days of their releases.
On Windows, the hardware and antivirus matrix is the boss fight — features are easy compared to running correctly on every GPU driver and under every real-time scanner.
No forced subscription — that's the style
The IDE is free with your own API keys, forever — you pay model providers directly, and at BYOK prices that's dramatically cheaper than fixed AI-IDE fees. The $10/month Pro plan exists for the things that genuinely need servers — hosted web search, semantic embeddings, OCR, MCP connections — and as a way to support the project if you believe in it. Subscribe when you need it, cancel when you don't, and the IDE keeps working either way. No pressure, no forced auto-renewal games. One person built this for the public; the business model has to respect the public too.
What's next
The roadmap runs on the same loop as the product: real user reports in, verified fixes out — every release note answers an actual email. If you want to pressure-test twelve months of one person's 14-hour days, download Cortex, hand it a real task, and tell me what breaks at [email protected]. Anonymous downloads mean I won't know you're there — so the emails are how I learn.