Fast, private, comment-aware code completion that runs locally.

Highlights

Inline and IntelliSense completion without an account or subscription.
Purely syntactic-statistical prediction: no LLM, neural runtime, or model server.
Comments are excluded from code context and training.
Variable-order n-grams with exact backoff instead of model-wide fuzzy scans.
Deterministic structural filtering for member, expression, and JSON positions.
Bounded project identifier trie and numeric adaptive n-gram context storage.
Bounded multi-token suggestions with Fast, Balanced, and Quality presets.
Focused automatic frequency control and shorter five-token suggestions.
Conservative cross-project learning from saved, error-free files. A pattern needs support from multiple projects and never outweighs the bundled model.
Receiver-aware API member patterns and useful numeric expressions.
Optional local project adaptation and free language-pack support.
Local diagnostics with no telemetry.

Supported profiles: C#, Java, JavaScript, TypeScript, JSON/JSONC, Python, JSX/TSX, Vue, and Razor.

Source, releases, and issue tracking: kevinski119/local-ngram-code-suggester.

Offline model demo

The following animation is a deterministic Pillow-rendered editor simulation. Its completion is queried from the bundled n-gram model; it is not an API response or a recording containing personal desktop data.

Offline n-gram completion demo

Local model diagnostics

Free language-pack manager

Run Local N-Gram: Show Model Diagnostics to inspect the active model and latency. Run Local N-Gram: Manage Language Packs to import a local pack or use a configured trusted catalog. No pack is downloaded without user action.

Privacy

Tokenization, completion, project context, and diagnostics stay on your machine. Network access is limited to user-initiated pack catalog/download operations or explicitly opted-in pack update checks.

Cross-project learning stays in VS Code global storage on this machine. It stores hashed project identities, caps each project's contribution, ignores files with reported errors, and requires multiple-project support before a pattern can be suggested. Learned library identifiers are namespaced by detected pip, npm, Java, or .NET imports and activate only when the same dependency is present. Run Local N-Gram: Clear Cross-Project Learning at any time to erase it.