Local AI Toolkit

Open-source project guide

Русская версия

Local AI Toolkit is Windows developer tooling under the MIT license. You can use, modify and redistribute the project’s code, including commercially, subject to that license. Dependencies and model weights retain their own licenses.

Choose a starting point

Your goal Start here
Install the current release Installation and downloads
Understand the system How it works and responsibility map
Connect a repository Command-line walkthrough
Check data handling Privacy and security policy
Understand token estimates Token comparison
Contribute a change Contributing and governance
Request help Support
Publish a release Publication checklist and code signing policy

Project boundaries

The launcher gives clients a stable executable path. MCP servers and CLIs share a durable broker; only that broker sends inference requests to the configured model backend. Code search combines embeddings with symbol/name matching, using a repository base index and worktree overlays. Repositories are connected explicitly.

flowchart LR
  clients["Codex / Claude Code / CLI"] --> launcher["Stable launcher"]
  launcher --> tools["CodeSearch / LocalLm"]
  tools --> index["Base index + worktree overlay"]
  tools --> broker["Shared broker and queue"]
  broker --> backend["Configured model backend"]

Local inference does not mean that returned excerpts never reach the cloud: the calling client receives them and controls what it forwards. A GPU policy is a resource constraint, not a privacy guarantee. See Privacy.

Documentation surfaces

Distribution and trust

Current releases use an ECDSA-signed manifest and package hashes. Authenticode onboarding with SignPath Foundation is a separate process; do not assume a Windows publisher signature until a release explicitly identifies it. GitHub attestations of uploaded assets do not, on their own, prove those assets were built by CI from the tagged source.

See code signing policy. No commercial support, availability guarantee or certificate approval is promised by this documentation.