VNBtec — the AI that proves your code compiles

VNBtec is a multi-Apps Cluster. You enter once, one account, one brain (VNB-Brain), and four sections open to you: Code, Security, Science, Social, each with its own apps. This extension is the Code section.

VNBtec is an AI coding assistant for VS Code. What makes it different is not the model — it is this rule:

The best-scoring answer that compiles wins. And if none compiles, it says so.

Every other assistant hands you code and hopes. VNBtec verifies it before it reaches your file, and it tells you the verdict instead of hiding it behind persuasive prose.

The ghost text, and the badge

Native ghost text. Suggestions appear as you type (Tab to accept), with debounce, cancellation, a time ceiling and a cache. If a suggestion would arrive too late to be useful, it is not shown at all.
Verified, or stated as unverified. Run VNBtec: Generate Code (verified) (Ctrl+Alt+V) and the result carries its verdict: a status-bar badge, a diagnostic pinned to the inserted range, and — when the server could not prove the code — a refusal to insert it until you say otherwise.

The result has three states, never two: proven (it compiles), refuted (nothing compiled), and nothing to verify (the answer was prose). The third one is never reported as the first.

Everything else, in one account

Chat — the VNBtec assistant, in the sidebar.
Explain / find bugs / refactor / write tests on any selection.
Math (step-by-step, verified with SymPy), Physics & Chemistry.
Templates, Talk (messaging) and the Hub.

Getting started

Get an API token at vnbtec.com → Settings → API Keys (a free account includes 500k tokens).
Open Settings (Ctrl+,) and search for vnbtec:
- vnbtec.apiToken — paste your token (starts with sk-vnbtec-)
- vnbtec.apiUrl — leave as https://api.vnbtec.com unless you self-host
Start typing in any file: the ghost text appears. Ctrl+Shift+L opens the chat.

Autocompletion can be toggled from the status bar, or with VNBtec: Toggle Autocompletion.

Settings

Setting	Default	What it does
`vnbtec.inline.enabled`	`true`	Suggestions as you type
`vnbtec.inline.debounceMs`	`350`	Quiet time before asking — lower feels faster, higher spends fewer tokens
`vnbtec.inline.timeoutMs`	`4000`	After this delay, no suggestion is shown at all
`vnbtec.inline.maxTokens`	`256`	Budget for one suggestion (the server caps it too)
`vnbtec.verification.blockUnverified`	`true`	Refuse to insert code the server could not verify

Commands

Command	Keybinding
`VNBtec: Open AI Chat`	`Ctrl+Shift+L`
`VNBtec: Inline Suggestion`	`Ctrl+Shift+I`
`VNBtec: Generate Code (verified)`	`Ctrl+Alt+V`
`VNBtec: Explain Selection` / `Find Bugs` / `Refactor` / `Write Tests`	context menu

About the numbers

We publish what we measure, with its date and its limits. On the frozen code-v1-12 task set, the delivered code compiled and defined what was asked in 12/12 cases (level 1, measured 14 October 2026). That is not "the computation is correct" — that requires sandboxed execution, which is the next level. The distinction is written everywhere we show the figure, including on the website.

Licence

See LICENSE.txt.