AJRouter Engine

What AJRouter is

A local Windows control plane for the bundled llama.cpp and audio.cpp runtimes. It runs and supervises them as workloads, gives you a dashboard at http://127.0.0.1:8077, and exposes an OpenAI-compatible text API for your tools to talk to. It is not a model and not a replacement for llama.cpp: it is the layer around them.

What's in the package (version 2026.09.25)

• AJRouter control plane and dashboard: start, stop and restart workloads, read logs, see status. The dashboard is loopback-only. No firewall rules are created or changed. Nothing phones home.
• Four workloads: text (llama.cpp router with an OpenAI-compatible API), TTS, ASR and music (audio.cpp). Audio workloads ship disabled until you configure a model for them.
• Model routing: one loaded model at a time by default, GPU exclusivity, idle unload after a TTL, per-profile settings in config\models.ini.
• Coding-agent workspace with project templates: FiveM resource, Python (global or venv), TypeScript, Vite, Vue + Vite. Needs Python 3.12 or newer on your machine, for this feature only.
• llama.cpp build 11104 (CUDA and CPU backends) with its tools: llama-server, llama-cli, llama-bench, llama-quantize, llama-tts, the multimodal CLIs and more.
• audio.cpp server, CLI and GGUF tool, with the CUDA 13 runtime libraries.
• Installer (install.ps1): installs to %LOCALAPPDATA%\AJRouter, adds it to your PATH, creates Start-menu and desktop shortcuts; a repair install keeps your config. Or run start-ajrouter.cmd in place without installing.

What's not included

• No model weights, model recommendations or download links. You choose your GGUF models, put them in storage\models, and add a profile to config\models.ini. Model licensing and hardware fit stay with you.
• No cloud account, no subscription, no telemetry.

Requirements

• Windows 10 or 11, 64-bit. Windows PowerShell 5.1 or PowerShell 7.
• An NVIDIA GPU with a driver that supports CUDA 13, for GPU acceleration.
• Python 3.12 or newer, only if you use the coding-agent workspace.
• About 1.8 GB of disk for the engine, plus your models. VRAM needs depend entirely on the models you choose.

Editions

Personal: personal and non-commercial local AI use.
Commercial: businesses, revenue-generating projects and paid client work.
Same engine in both; the intended use is what changes.

Third-party components

Bundles llama.cpp (MIT, © The ggml authors), audio.cpp (Apache 2.0, © 2026 ShugoAI LLC), the LLVM OpenMP runtime (Apache 2.0 with LLVM exceptions) and NVIDIA CUDA runtime libraries. Licence texts are included in the package.

Support

Setup and defect support is included: contact me through ajthe.dev or Gumroad if something does not work as described here.