local-llm-advisor

A Claude Code skill that tells you which open-source models you can actually run on your machine, and what to use each one for.

It probes your hardware, works out what you'd realistically use a local model for, then researches what's current — and reports back with real file sizes and runnable commands.

Install

git clone https://github.com/foundanand/local-llm-advisor.git ~/.claude/skills/local-llm-advisor

Then in Claude Code:

/local-llm-advisor                 # full report
/local-llm-advisor coding          # narrow to one use case
/local-llm-advisor --install       # also pull the top pick

How it works

1. Probe — scripts/probe.sh reads chip, cores, RAM/VRAM, free disk, installed runtimes (Ollama, LM Studio, MLX, llama.cpp), already-downloaded models, and which AI tool configs you have.
2. Infer workloads — from your projects and configs: coding, vision, embeddings, bulk text, privacy-bound work. Local models are framed as complements to whatever cloud tools you already use.
3. Research — builds the candidate list from Hugging Face sorted by recency, then verifies sizes against the actual repo file lists.
4. Report — per workload: model, release date, real size, speed class, and the exact command for a runtime you already have.

Design notes

Three rules that exist because the first version got them wrong:

• Never search "best <category> <year>" to build the list. Those articles rank on backlink age, so they structurally cannot contain a model released this month. Start from HF sort=trending / sort=createdAt instead; use name-specific searches only to research candidates you've already found.
• Never compute size from parameter count. Look up the actual file. A hardcoded Q4 ≈ 0.6 GB/B table makes sub-4-bit builds arithmetically undiscoverable — which is how a 27B model that fits in 7 GB gets ruled out as "too big."
• Hugging Face is the universe; Ollama's library is a lagging subset. New weights land on HF within days and reach curated registries much later, if ever. Anything on HF runs via ollama run hf.co/<repo>, llama-server -hf, or mlx_lm.

What the probe reads

Local only, read-only, nothing is transmitted. It lists paths of AI tool configs (~/.claude, ~/.gemini, etc.) and your recent project names to infer workloads — it does not read project contents. Read scripts/probe.sh before running it if you'd rather check.

License

MIT