One workaround if rate limits keep biting: use a proxy that can failover to alternative providers when one hits limits. I built NadirClaw for this - it classifies prompts and routes to the cheapest model that can handle the task, but also switches providers when one is unavailable. Doesn't help if you specifically need Claude's quality for every task, but for the "pointer on localization" type questions it could route those to a cheaper model entirely and save your Claude quota for the hard stuff.
Nice work on the observability side. We built something complementary - NadirClaw routes requests across multiple providers based on cost/performance tradeoffs. Your Prism setup shows which departments are spending, then a routing layer like NadirClaw can cut that spend 40-70% by sending simple prompts to cheaper models when appropriate. The two would work well together - tracking + optimization.
Interesting. If you're already proxying LLM requests for security, adding cost-optimized routing is a natural next step. NadirClaw does the routing side (classifies prompts, routes to cheaper models). Could see these stacking well. https://github.com/doramirdor/NadirClaw
Cool project. If anyone wants something lighter that works as a drop-in OpenAI proxy (no code changes needed), I built NadirClaw for exactly this. It classifies prompts in ~10ms and routes to cheap/local models automatically. Works with Claude Code, Cursor, aider out of the box. https://github.com/doramirdor/NadirClaw (author)
I use Claude and Codex heavily for coding, and I kept burning through my quota halfway through the week. When I looked at my logs, most of my prompts were things like "summarize this" or "reformat this JSON." Stuff any small model handles fine.
NadirClaw is a Python proxy that classifies each prompt in ~10ms and routes simple ones to Gemini Flash, Ollama, or whatever cheap model you want. Only complex prompts hit your premium API. It's OpenAI-compatible, so you just point your existing tools at it.
In practice I went from burning through my Claude quota in 2 days to having it last the full week. Costs dropped around 60%.
pip install nadirclaw
Still early. The classifier is simple (token count + pattern matching + optional embeddings). Curious what breaks first.