The catalogue is the input; the value is the arithmetic on top — VRAM at a given quant against real GPU tiers, and an API-versus-self-hosted cost table in EUR. One caveat matters for the cost side: the API and hardware prices are hardcoded as of 2026-05 and the README asks for them to be verified semestrially, so treat the VRAM answer as durable and the euro figures as a snapshot.
An MCP server over the quelllm.fr catalogue of 190+ open-weights LLMs. It turns the catalogue into six tools a client can call: filter and search models, read one model's full record, compare two of them, size the VRAM at a chosen quantisation, and price API usage against self-hosting.
- Models filtered by origin, family or maximum parameter count — `list_models`
- One model's full record: params, VRAM per quant, context window, family, tags, license and URLs — `get_model`
- A side-by-side comparison of two models with a verdict — `compare`
- VRAM in GB at a chosen quantisation, plus the GPU and Mac tiers that fit — `estimate_vram`
- A monthly cost table in EUR comparing API providers against self-hosted hardware — `estimate_cost`
- Fuzzy search across name, family, tag and author — `search_models`
No key. Install from source with `pip install git+https://github.com/MGM-FALCON/quelllm-mcp.git`, or run it without installing via `uvx --from git+https://github.com/MGM-FALCON/quelllm-mcp.git quelllm-mcp` — so `uv` or `pip` plus network access is the whole setup. Catalogue data comes from the public quelllm.fr API (CC BY 4.0, no key) and is cached locally for 1h. The version is 1.1.0.
One command — pip install git+https://github.com/MGM-FALCON/quelllm-mcp.git