1 The short answer
Cursor is a subscription AI editor with bundled inference. Cortex is a free agentic editor you point at your own API keys. That single difference drives almost everything else below: who sets your model list, how your bill behaves under heavy agent use, and what happens to your workflow if the vendor changes pricing or ownership.
Cursor is the more mature product. It has a larger ecosystem, deeper tab-completion tuning, and years of polish behind its editor experience. Cortex is younger and narrower, and it trades that breadth for two things Cursor's model cannot offer: no markup on your tokens, and no credit meter between you and the model.
You want one managed bill
You would rather pay a flat subscription than track per-token spend, you want hosted models with no key setup, and you value the largest extension and community ecosystem in the category.
You want to own the model choice
You already have provider keys, you run heavy agent loops and want list-price tokens, you need local or self-hosted models, or you do not want your editor's roadmap decided by whoever acquires it next.
2 Side-by-side table
Everything here is checkable. Where a Cursor detail can change without notice, the row says so rather than asserting a number we cannot keep current.
| Dimension | Cortex AI IDE | Cursor |
|---|---|---|
| Licensing model | Free with BYOK. Pro is $5/month or $40/year per seat for semantic search, web search, MCP and voice input. | Subscription tiers. Agent usage is metered in credits rather than billed at provider list price. |
| Who you pay for tokens | Your model provider, directly. No markup and no resale of your usage. | Cursor, as part of the subscription. Bundled inference is the product. |
| Model access | Any provider you add a key for, plus any OpenAI-compatible endpoint via Cortex Connect: Ollama, vLLM, LM Studio, Groq, Together, or your own gateway. | Cursor's hosted model list. A custom OpenAI-compatible key option exists, but the agent is built around Cursor's own inference. |
| Local and air-gapped models | Supported. Point Connect at a local endpoint and no prompt leaves the machine except to that endpoint. | Not the design centre. Offline or fully self-hosted model use is limited. |
| Editor architecture | Native desktop app with Monaco as the editor surface, a real file explorer, and an integrated terminal. | A fork of VS Code, so it inherits the Electron shell and the VS Code extension ecosystem. |
| Platforms | Windows 10/11 and Linux as released builds. macOS via the official maintainer-authorized build guide. | Windows, macOS and Linux. |
| Agent tooling | 13 core agent tools: read, write, edit, bash, glob, grep, semantic search, LS, todo list, ask-user, web search, web fetch, and a reviewed loop. See the tools reference. | Agent mode with multi-file edits, terminal commands and its own tool set. Not published as a fixed public count. |
| Verification of agent work | Cross-file impact detection runs on every Python edit and warns in the same turn when a renamed or removed symbol is still imported elsewhere. | Relies on the model and on you running the tests. No equivalent static cross-file gate is documented. |
| Where your keys live | On your machine, in the OS keychain. Prompts and code go to your provider, never through our servers. | Managed by Cursor as part of the hosted service. |
| Source availability | Desktop source published under Apache-2.0, with the macOS port authorized in writing. | Closed source. |
| Ownership | Built by Pasons Group, independent. | Anysphere, under a $60 billion all-stock acquisition agreement from SpaceX announced June 16, 2026. |
| MCP servers | Supported on Pro, configured from Settings. See the MCP docs. | Supported. |
Competitor pricing, credit allowances and model lists change often and without notice. Treat the Cursor column as the shape of the product, not a quote of today's price page, and check cursor.com before you decide on cost alone.
3 What each actually costs
The sticker price is not the real number for either tool. What matters is what a month of agent work costs once you are running multi-file edits, terminal commands and test loops all day.
With Cortex you pay your provider's list price per token. There is no markup, so the bill scales with the work and drops immediately when you route cheap models to cheap tasks. The IDE itself is free; Pro is $5/month or $40/year per seat and adds semantic code search, web search, MCP servers and voice input.
With Cursor you pay a subscription and your agent usage draws down a credit allowance. That is predictable when you are inside the allowance and expensive when you are not. In June 2026 Cursor moved from unlimited agent use to a 1,500-credit monthly allowance, which is the kind of change a bundled-inference business can make at any time.
We published the real per-token numbers, including what one agentic task costs on each supported provider, in Cheapest Coding Agent Models: Real BYOK Prices. Read that before you compare monthly fees, because the fee is usually the smaller half of the total.
4 Models and keys
Cursor decides which models you can use and what they cost you. That is a genuine convenience: no keys to create, no rate limits to manage, no provider console to log into. It is also a dependency. When the model list changes, your options change with it.
Cortex takes the opposite trade. You add keys for the providers you want, and the model list is yours. Cortex Connect extends that to any OpenAI-compatible endpoint, so a local Ollama instance, a vLLM server on your own hardware, or a company gateway all work the same way as a hosted provider. Providers and models can be added one by one or imported by CSV.
If you are new to the model, What Is BYOK in AI Coding? explains the mechanics and the security side of storing keys locally.
5 Editor architecture
Cursor is a VS Code fork. You get the extension ecosystem, the keybindings and the familiarity immediately, and you inherit the Electron shell. For most people that is the right call, and switching cost is near zero if you already use VS Code.
Cortex is a native desktop application with Monaco as its editor surface, a real file explorer, integrated terminal, and search. It does not run the VS Code extension ecosystem. What it keeps, deliberately, is the code itself: the file tree, search and editor stay one click from the chat instead of being collapsed behind it. The reasoning is in Why Cortex Never Hides Your Code.
On platforms, Cursor ships Windows, macOS and Linux builds. Cortex ships Windows and Linux builds, and publishes an authorized macOS port guide that builds it from source on Apple silicon or Intel in roughly 30 to 45 minutes. If you need a signed, notarized Mac installer today, that is a real point in Cursor's favour.
6 The agent loop
Both tools run an agentic loop: read the codebase, plan, edit across files, run commands, check the result. The difference is in how much of that loop you can see and interrupt.
Cortex works in three modes. Agent plans and executes with a permission card before any write, edit or delete. Ask answers without touching files. Plan produces the plan and stops for your decision. Every tool call renders as a card in the chat, so you can watch the work rather than wait for a summary of it. Agentic Loop Mode adds a formal cycle where an independent reviewer model audits the changes for missing tests, secrets in code and incomplete work before the run is allowed to finish.
The verification step is where the two diverge most. After every Python edit, Cortex's cross-file impact detector parses the change and warns in the same turn, with exact file and line locations, if a function the agent renamed or removed is still imported somewhere else. That catches the class of breakage you would otherwise find as an ImportError at runtime. The full loop is described in What Is Agentic Coding?, and why it matters for trust in Why Only 29% of Developers Trust AI Code.
7 Privacy and ownership
Cortex is privacy-first by architecture rather than by policy. Your API keys are stored in the operating system keychain on your machine. Your prompts and your code go directly to the provider you configured and never pass through our servers. There is no telemetry pipeline carrying your source.
Cursor is a hosted service, so its inference path necessarily involves its own infrastructure. That is not a criticism, it is the design, and it is what makes the zero-setup experience possible. It does mean your trust decision is about a company rather than about a network path you can inspect.
The ownership question got sharper in 2026. SpaceX agreed to buy Anysphere for $60 billion in June, after Google acqui-hired Windsurf's founders and Cognition folded the rest of Windsurf into Devin Desktop, retiring Cascade on a hard deadline. The standalone AI IDE company is disappearing. Cortex is independent, and its desktop source is published under Apache-2.0, so the product cannot be retired by an acquirer's roadmap decision. The full story is in SpaceX Bought Cursor for $60 Billion and Windsurf Is Now Devin Desktop.
8 Which should you pick?
Neither choice is irreversible, and you do not have to make it exclusively. Both editors open the same project folder, so the practical path is to keep shipping from the tool you already know and add the other where it is stronger.
- Stay on Cursor if the managed subscription is worth more to you than per-token control, if you depend on the VS Code extension ecosystem, or if you need a notarized macOS build today.
- Move to Cortex if you already hold provider keys, if heavy agent use is pushing you past a credit allowance, if you need local or self-hosted models, or if you want the editor's future to be independent of an acquirer.
- Run both if you want a hedge. Keep your keys in your own name and keep a local agent that does not depend on a conglomerate login. That is the position we argue for in the acquisition piece, and it is the reason Cortex is free with BYOK.
If you are still comparing editors more broadly, How to Choose an AI IDE in 2026 walks through native against Electron, agentic loop against autocomplete, and BYOK against bundled inference, and Best Free AI Coding IDEs in 2026 separates genuinely free from free-tier.