After more than a year of nights and weekends, a developer known as TangySword has released Jenny, an MIT-licensed Electron desktop app for running LLMs locally. It is not just a chat interface: it combines tool calling, rollback, and an IDE, and it starts from an explicit belief that frontier model access is currently subsidized by private equity and venture capital, and that when those subsidies dry up the service will degrade. From there comes the decision to build a self-hosted harness that does not depend on cloud endpoints.

The project arrives in a crowded segment — LM Studio, Open WebUI, Unsloth Desktop — but it leans on a particular balance between productivity and local safety. Destructive shell commands require explicit approval; file edits are checkpointed. For users of small models, often fast but error-prone, this is not an accessory: it reduces the cost of a bad output and makes automation on files and calendars more practical. Jenny also exposes diagnostics and logs so users can reconstruct what the local runtime did.

Technically, the app accepts llama.cpp, vLLM, or any OpenAI-compatible local endpoint, plus GGUF through the managed llama-server (for MTP). The developer reports solid performance on a GeForce RTX 5070 Ti with a 9-billion-parameter model, ornith1.5:9b, for tool calling and speed. Windows is the reference platform; macOS is untested and Linux is not available. That choice is consistent with a desktop-first priority on a single workstation, but it limits headless server adoption in enterprise settings.

The structural point is not the individual app but the signal it sends. When the quality of commercial APIs depends on outside subsidies, the local route stops being an ideological niche and becomes a way to control TCO and data. Jenny addresses this at the individual user level: no network calls beyond the local runtime, a permissive license, and private telemetry for debugging only. The maintenance risk remains: this is a solo project, with a day job and a private life to protect, and the open code is more of an invitation than a guarantee. But for anyone looking at on-premise deployment, tools like this show that the bottleneck is not raw compute, but the ability to govern the side effects of a model that gets things wrong. For those evaluating these trade-offs, AI-RADAR offers analytical frameworks at /llm-onpremise; the point here is different: local execution demand is producing increasingly complete interfaces, even when built by a single person.