Claude Code local models hit OOM

My read: Claude Code looks strong on hosted models, but local-model support still feels brittle enough that token budgeting becomes the real product.

• –The failure mode is likely context growth plus output limits, not just raw model size, so VRAM pressure can spike fast.
• –The thread suggests the UX around `max token` versus `max budget token` is not obvious for new users.
• –Local inference is attractive for privacy and cost control, but coding agents are especially sensitive to long prompts, tool traces, and retries.
• –If Claude Code is optimized around Anthropic-hosted models, swapping in a local backend probably needs more manual guardrails than most users expect.