abhiindigloo/model-finder

Run the user's own examples against several AI models (Claude, Gemini, OpenAI, or any OpenAI-compatible endpoint including local Ollama or vLLM) and measure pass rate, latency, and billed cost to find the cheapest model that meets their quality bar. Use whenever the user wants to test, benchmark, evaluate, or prove which model works for their task, asks "is the cheaper model good enough", wants evidence before switching models, or has a shortlist from find-model to validate, even if they don't say "eval". Generates a config, runs a standard-library Python script with the user's keys from environment variables, and writes results.json plus a summary.

Compare and recommend AI models across providers (Anthropic Claude, Google Gemini, OpenAI, Bedrock, Vertex AI, Azure, Mistral, open-weight and local models) for a specific task, budget, and deployment route. Use whenever the user asks which model to use, whether a cheaper model is good enough, how models compare on quality, cost, speed, context, privacy, or availability, which model suits a coding subtask in their IDE or agent (planning, edits, review, tests, autocomplete), or mentions model finder (or "modal finder"), even if they never say "compare". For running a real evaluation use eval-models; for fallback or routing configs use route-models; for local hardware memory use size-local-model.

Generate a cheapest-first model routing setup with fallbacks and quality escalation, from eval-models results or a user-chosen model order. Produces a standalone Python router (escalates when a reply errors or fails format guards), a routing.json, and a LiteLLM proxy config (error-based fallbacks). Use whenever the user wants model routing, fallbacks, failover, escalation to a stronger model, cost-saving cascades, "use the cheap model unless it fails", or a LiteLLM or gateway config across Claude, Gemini, OpenAI, or local models, even if they don't say "routing". Never deploys or applies anything.

Estimate the GPU or unified memory needed to run an open-weight model locally, covering weights at a given precision or quantization plus KV cache for a chosen context length and concurrency, and compare it with available memory. Use whenever the user asks whether a model fits their GPU, Mac, or server, how much VRAM or RAM a model needs, what quantization to use, how context length or batch size affects memory, or how to size hardware for Ollama, llama.cpp, vLLM, or LM Studio, even if they don't say "memory". Gives estimates with stated assumptions, never a guarantee of fit.