Skip to main content
Glama

quelllm-mcp

MCP server exposing the quelllm.fr catalog of 190+ open-weights LLMs via Model Context Protocol tools. Use it from Claude Code, Cursor, Continue, or any MCP-compatible client to query models, compare them, estimate VRAM, and compute API vs self-hosted cost.

Tools exposed

Tool

Description

list_models(filter_origin?, filter_family?, max_params_b?)

List models with filters (origin code, family, max params in B)

get_model(model_id)

Full record for one model (params, vram per quant, context window, family, tags, license, URLs)

compare(model_a_id, model_b_id)

Side-by-side comparison with verdict

estimate_vram(model_id, quant)

VRAM in GB at chosen quant + recommended GPU/Mac tiers

estimate_cost(input_tokens_per_month, output_tokens_per_month, ...)

Cost in EUR — full table API providers vs self-hosted hardware OR a specific id

search_models(query, limit?)

Fuzzy search by name, family, tag, author

Related MCP server: mcp-turboquant

Install

Install from source (not yet on PyPI) :

pip install git+https://github.com/MGM-FALCON/quelllm-mcp.git

Or run without installing, using uv :

uvx --from git+https://github.com/MGM-FALCON/quelllm-mcp.git quelllm-mcp

For local development :

git clone https://github.com/MGM-FALCON/quelllm-mcp.git
cd quelllm-mcp
pip install -e .

Use with Claude Code

Add to ~/.claude.json or a project's .mcp.json. If you installed with pip :

{
  "mcpServers": {
    "quelllm": {
      "command": "quelllm-mcp"
    }
  }
}

Or zero-install with uvx :

{
  "mcpServers": {
    "quelllm": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/MGM-FALCON/quelllm-mcp.git", "quelllm-mcp"]
    }
  }
}

Use with Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) :

{
  "mcpServers": {
    "quelllm": {
      "command": "quelllm-mcp"
    }
  }
}

Use with Cursor / Continue / Cline

Most MCP clients accept the same JSON config :

{
  "command": "quelllm-mcp"
}

Example queries (from your client)

> Quels LLM Mistral peuvent tourner sur RTX 5070 Ti 16GB ?
→ list_models(filter_family='Mistral', max_params_b=24)
→ estimate_vram('mistral-small-24b', 'q4')

> Compare Llama 3.3 70B vs Qwen 2.5 32B
→ compare('llama33-70b', 'qwen25-32b')

> J'utilise 10M tokens input + 2.5M output / mois. Combien je paye chez OpenAI vs DeepSeek ?
→ estimate_cost(10_000_000, 2_500_000)

Data source

All data pulled from quelllm.fr/api/ (CC BY 4.0, no key, CORS-enabled). Cached locally for 1h to avoid rate-limiting.

API pricing data (GPT-5, Claude Opus 4.7, Gemini 2.5, DeepSeek, Mistral) and hardware pricing (RTX 50-series, Mac M4) are hardcoded as of 2026-05 — verify semestrially.

Model-selection guides by language

The MCP server exposes the same catalog to any compatible client. These companion guides explain model selection, memory trade-offs and repeatable evaluation tasks for different languages. They are source-based guides, not new hardware benchmarks.

License

MIT — see LICENSE.

Contributing

Source : https://github.com/MGM-FALCON/quelllm-mcp Issues + PRs welcome. Particularly :

  • API pricing updates (semestrial)

  • Hardware additions (new GPUs, Mac Mx series)

  • New tools (e.g. find_alternatives_to(model_id), recommend_gpu(budget_eur))

Tests

A pytest smoke suite lives under tests/. It covers all 6 tools and the v1.1.0 output invariants, never touches the network (local fixture + mocked httpx), and stubs the mcp SDK when it isn't importable — so it also runs on Python 3.9.

pip install -e ".[test]"
pytest

Author

Mohamed Meguedmi — LinkedIn · Hugging Face Founder of La Gazette IA and QuelLLM.fr.

Available Tools

6 tools
compareA

Compare two LLMs side-by-side.

Args: model_a_id: first model ID model_b_id: second model ID

Returns: comparison dict with VRAM, params, license, context window, family, origin for both models, plus a verdict

Inputs: model_a_id (required), model_b_id (required).

estimate_costA

Estimate monthly cost for LLM usage — API providers OR self-hosted hardware.

Args: input_tokens_per_month: total input tokens per month output_tokens_per_month: total output tokens per month model_or_hardware: specific model/hw ID (e.g. 'gpt-5', 'rtx-5090'). If None, returns full comparison table. amort_months: amortization period for hardware purchase, default 24 months electricity_eur_per_kwh: electricity rate, default 0.20 EUR/kWh (FR 2026 average) hours_active_per_day: hours of active inference per day, default 4 caching_hit_rate: 0..0.9 fraction of input tokens hitting prompt cache (Anthropic 90% off / Gemini 75% off). Default 0. batch_mode: True applies 50% off input+output for batchable APIs (24h delay). Default False. off_peak: True applies DeepSeek off-peak discount (50% V3.5 / 75% R1). Default False. pc_always_on: True = idle (24-hours_active) × idleW. False = PC off out of active hours (0W idle). Default True. vat_recoverable: True = subtract 16.67% VAT from hardware purchase price (HT pour pro). Default False.

Returns: cost estimate in EUR (with USD-to-EUR conversion ~0.92 for APIs)

Inputs: input_tokens_per_month (required), output_tokens_per_month (required), off_peak, batch_mode, +7 more.

estimate_vramA

Estimate VRAM required to run a model at a given quantization.

Args: model_id: the model ID quant: quantization level — one of 'q4', 'q5', 'q8', 'fp16'. Default 'q4'.

Returns: dict with vram_gb (estimate including context overhead) + recommended GPU tiers

Inputs: model_id (required), quant.

get_modelA

Get full details for a single model.

Args: model_id: the model ID (e.g. 'mistral-7b-instruct', 'qwen3-8b')

Returns: full model record with params, vram per quant, context window, family, tags, license, and URLs.

Inputs: model_id (required).

list_modelsA

List open-weights LLMs from quelllm.fr catalog (190+ models).

Args: filter_origin: filter by author origin code, e.g. 'fr', 'us', 'cn' filter_family: filter by model family, e.g. 'Mistral', 'Qwen', 'Llama' max_params_b: maximum number of params in billions (e.g. 32 for ≤32B models)

Returns: dict with keys: count, models (list of {id, name, author, params, family, license, vram_q4_gb})

Inputs: max_params_b, filter_family, filter_origin.

search_modelsA

Fuzzy search models by name, family, tag, or author.

Args: query: search string limit: max number of results, default 10

Returns: dict with count and matching models (sorted by relevance)

Inputs: query (required), limit.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose: listing models, searching, getting details, comparing, estimating cost, and estimating VRAM. No significant overlap or ambiguity.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern in snake_case (e.g., list_models, estimate_cost), making it easy to predict functionality from the name.

    Tool Count5/5

    With 6 tools, the count is appropriate for the domain of LLM discovery and analysis. Each tool serves a clear purpose without unnecessary bloat or deficiency.

    Completeness4/5

    The toolset covers core operations: listing, searching, detail retrieval, comparison, cost estimation, and VRAM estimation. Minor gaps exist, such as direct pricing for specific models, but the overall workflow is well-supported.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      D
      maintenance
      An MCP server that enables users to query, compare, and synthesize responses from multiple local and cloud LLMs simultaneously using existing subscriptions. It provides tools for parallel model evaluation, consensus polling with an LLM-as-judge, and response synthesis across different model providers.
      8
      59 npm
      15
      MIT
    • F
      license
      A
      quality
      D
      maintenance
      An MCP server for discovering and querying over 300 AI models available on OpenRouter. It enables users to list, search, filter, compare, and get detailed information about models with pricing, context limits, and capabilities.
      5
      1
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      InferBench's MCP server lets coding agents run, serve and benchmark local LLMs (text + image, llama.cpp + Stable Diffusion) on your own hardware on demand — measuring real tokens/sec and picking the optimal quant for your GPU from a 124-model catalog. Local-first, no cloud required.
      2
      MIT