Local AI Toolkit

Dependencies

Contents · Русская версия

The installer checks each one. Git, Ollama, the GitHub CLI, the .NET SDK, Node.js and Python it offers to install through WinGet, and the two SCIP indexers through npm once Node.js is there; where WinGet is missing, it names each one’s official download page instead.

Without these it does not work at all:

  Why
Windows Windows 10 or 11 on x64; the installer reports anything else as unsupported
Git a repository is identified by its git directory, and the hooks live there
Ollama every model runs on it; Local AI only ever reaches it through the broker
A GPU the model fits in under the default policy, RequireFullVram, a model that does not fit entirely in video memory is refused rather than run slowly. The policy can be relaxed — see below

Needed per language — search works without them, but coarser:

  Why
.NET SDK 10 the C# semantic index behind precise navigation is built by loading the solution through the installed SDK. Cutting C# into types and members does not need it: Roslyn does that inside Local AI
Node.js 20 + SCIP TypeScript declaration-level chunking for TypeScript and JavaScript
Python 3.10+ + SCIP Python declaration-level chunking for Python

Optional:

  Why
GitHub CLI used only when the anonymous HTTPS download of a release fails — a private repository, or a release host that is unreachable while the API is not. The installer tries HTTPS without an account first, and installing from a folder needs neither

The residency policy can be relaxed deliberately: localai policy set --residency AllowPartialOffload or AllowCpu. A model that then runs partly or wholly outside video memory is marked [partial_offload] or [cpu] beside its name in the call’s report line.


← How it works · Contents · Installing →