Returning the five factors alongside the total is the design choice that makes the ranking usable — you can see whether a model is at the top because it is genuinely strong or because it was released last week. The hardware filters are described by the project itself as heuristic rather than verified compatibility, which is the right caveat to carry into any 'will this fit in 8GB' answer: treat the VRAM estimate as a shortlist, not a guarantee.
A hosted MCP server over the Free2AITools registry, a cross-source catalog of AI models, datasets, papers, tools and benchmarks. Entries are scored by a composite index built from five factors — Semantic, Authority, Popularity, Recency and Quality — and every API response carries the full factor breakdown rather than only the total.
- The catalog searched across sources including HuggingFace, GitHub, ArXiv, Ollama, Replicate and Civitai — `free2aitools_search`
- Results ranked by the composite index — `free2aitools_rank`
- The reasoning behind a score, factor by factor — `free2aitools_explain`
- Model selection under constraints such as a VRAM ceiling and a task category, returning ranked entries with parameters, VRAM estimates and license — `free2aitools_select_model`
- Two or more entries compared side by side with their factor decomposition — `free2aitools_compare`
- A factual factor summary rather than a verdict: the response reports what the index says and leaves the fit judgement to the caller
Nothing — no auth required and free to use. Add `"url": "https://free2aitools.com/api/mcp"` to your client's server list and it works in Claude, Cursor, Windsurf or any MCP-compatible client. The same catalog is also reachable as plain HTTP endpoints (`/api/v1/search`, `/api/v1/select`, `/api/v1/compare`) and as Parquet exports for offline analysis with DuckDB, Pandas or Spark. The repository package is `ai-nexus` (2.1.0), MIT licensed.
One command — npx -y mcp-remote https://free2aitools.com/api/mcp