Privacy-first coding agents go self-hosted

This is absolutely doable; the real question is whether you want to own the ops burden that comes with a private agent stack. The ecosystem already supports most of the plumbing, so the bottleneck is less compatibility and more model quality, latency, and day-to-day maintenance.

• –Cline already supports local models via Ollama or LM Studio, which makes it a natural fit for a private workflow: https://docs.cline.bot/getting-started/authorizing-with-cline
• –vLLM exposes an OpenAI-compatible server, so the proposed “self-hosted API, editor client on top” architecture is straightforward: https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html
• –Claude Code can also be routed through gateways and MCP tools, so the client layer is flexible if you want to swap providers later: https://docs.anthropic.com/en/docs/claude-code/llm-gateway and https://docs.anthropic.com/en/docs/claude-code/mcp
• –With 150-200 GB VRAM, the sweet spot is usually a strong open-weight coding model plus a faster smaller model for low-latency tasks; you’re buying privacy and control, not guaranteed frontier-level reliability
• –Vast’s hourly model fits bursty coding work well, but you inherit boot time, IP churn, auth management, and monitoring, which are the hidden cost of “privacy first”