Local Motion Brings Local LLM Agents to Cursor and VS Code

Local Motion is a VS Code/Cursor extension that gives developers a single switch to turn a local coding-model agent on or off on Mac. It profiles the machine's architecture, memory, and disk space, picks a compatible GGUF model, sizes a memory budget at 80% of unified memory pegged to Apple Silicon's Metal wired-memory ceiling, and fits the context window to that budget. The default setup installs a two-model pack — a larger Qwen3-Coder-30B-A3B and a lighter Gemma 4 26B A4B — running llama-server via Homebrew and registering the model with VS Code's Language Model API.

The most notable engineering detail is how it handles Cursor. Because Cursor's backend outright rejects requests to private network addresses like 127.0.0.1, Local Motion must front its local proxy with a public HTTPS tunnel via cloudflared before Cursor chat can reach it. That means prompts pass through both Cursor's servers and the tunnel — making the setup 'local compute' rather than fully private — whereas the vscode.lm provider in real VS Code runs entirely locally without any tunnel.

Beyond that tradeoff, the project focuses on practical memory management: instantly freeing RAM when idle, keeping multiple downloaded models on disk to switch between on demand, and falling back to alternate ports when needed. These details matter for engineers who want to bring local LLMs into their daily coding workflow while understanding exactly where privacy guarantees end.