Rate limits and subscriptions push people toward owning their stack

The thread 'If you don't run your own LLM proxy, how do you know what's going on?' got people nodding. One commenter is setting up LiteLLM on servers right now because they can't orchestrate agents in sandboxes without virtual keys that keep provider API keys out of reach. Another built their own agent harness for full observability. Meanwhile an older but still circulating thread has Anthropic cutting Claude Code subscriptions off from OpenClaw, with users complaining they hit rate limits within 1:45 on afternoons and that 'outsized strain' just means using what you paid for. A separate thread asks whether anyone has replaced Claude or GPT with a local model, and the answers are real: Qwen 3.6 27B dense judged roughly Haiku 4.5 level, DeepSeek V4 Flash on two RTX Pro 6000 cards at 160 tok/s, while Gemma 4 on an M4 was too slow.

The through-line: vendor-controlled access is a risk people now price in. A proxy gives you observability and key control, and a local model gives you a floor if a provider changes terms.

Nuance: local is not there for everyone. Throughput and missing enterprise tooling for model selection are the stated blockers, so most people are hedging, not switching.