io.github.TheBaronofAI/vetted-consumer

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Will a local LLM run on your hardware? GGUF quant, buy-vs-rent-vs-API cost, used-GPU prices.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "vetted-consumer": {
      "url": "https://vettedconsumer.com/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (9)

• can_i_run_it

  Will a given local LLM run on given hardware? Returns fit, the best quant that fits, theoretical tok/s, and real owner-measured tok/s where available.

  Parameters:

  • model — string — Model name, e.g. 'Llama 70B', 'gpt-oss-120B', 'Qwen 32B'. Use list_models to see known names.
  • mxfp4 — boolean — True if the model ships natively in MXFP4 (e.g. gpt-oss)
  • context — number — Context window in tokens (default 8192)
  • total_b — number — For an unlisted model: total parameters in billions
  • unified — boolean — True for unified-memory machines (Macs, Strix Halo, CPU+RAM)
  • vram_gb — number — For custom hardware: VRAM or unified memory in GB
  • active_b — number — For an unlisted model: active params in billions (= total for dense, less for MoE)
  • hardware — string — Hardware name/id, e.g. 'rtx-3090', 'Mac 128GB', 'Strix Halo'. Use list_hardware to see known ones.
  • kv_precision — string — KV cache precision (default f16)
  • bandwidth_gbps — number — For custom hardware: memory bandwidth in GB/s

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• recommend_quant

  Which GGUF quantization to download for a model on given hardware: the full quant ladder with file size, max context, and tok/s for each, plus the recommended pick.

  Parameters:

  • model — string — Model name, e.g. 'Llama 70B', 'gpt-oss-120B', 'Qwen 32B'. Use list_models to see known names.
  • mxfp4 — boolean — True if the model ships natively in MXFP4 (e.g. gpt-oss)
  • context — number — Context window in tokens (default 8192)
  • total_b — number — For an unlisted model: total parameters in billions
  • unified — boolean — True for unified-memory machines (Macs, Strix Halo, CPU+RAM)
  • vram_gb — number — For custom hardware: VRAM or unified memory in GB
  • active_b — number — For an unlisted model: active params in billions (= total for dense, less for MoE)
  • hardware — string — Hardware name/id, e.g. 'rtx-3090', 'Mac 128GB', 'Strix Halo'. Use list_hardware to see known ones.
  • kv_precision — string — KV cache precision (default f16)
  • bandwidth_gbps — number — For custom hardware: memory bandwidth in GB/s

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• cheapest_hardware_for_model

  The cheapest catalogued, buyable machine that runs a given model at Q4 with the requested context.

  Parameters:

  • model — string — Model name, e.g. 'Llama 70B', 'gpt-oss-120B', 'Qwen 32B'. Use list_models to see known names.
  • mxfp4 — boolean — True if the model ships natively in MXFP4 (e.g. gpt-oss)
  • context — number — Context window in tokens (default 8192)
  • total_b — number — For an unlisted model: total parameters in billions
  • active_b — number — For an unlisted model: active params in billions (= total for dense, less for MoE)

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• list_models

  List the local LLM model classes the tools know about (params, dense/MoE, native context).

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• list_hardware

  List the machines the tools know about (memory, bandwidth, price, buy link).

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• cost_compare

  Buy vs rent vs API cost to run a model locally: monthly/1y/3y totals, break-even months, and the energy cost per 1M tokens. Same math as /cost-calculator/.

  Parameters:

  • api — number — API $/million tokens (default 1.0)
  • kwh — number — Electricity $/kWh (default 0.16)
  • rent — number — Cloud GPU $/hour (default 0.59)
  • hours — number — Active hours per day (default 3)
  • tdp_w — number — For custom hardware: board power draw in watts
  • tokens — number — Tokens generated per day, for the API comparison (default 300000)
  • hardware — string — Catalogued hardware name/id (see list_hardware), e.g. 'rtx-3090-used'
  • price_usd — number — For custom hardware: price in USD

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• recommend_hardware

  Ranked list of catalogued, buyable machines that run a model at the requested context, cheapest first, with an optional budget cap.

  Parameters:

  • model — string — Model name, e.g. 'Llama 70B', 'gpt-oss-120B', 'Qwen 32B'. Use list_models to see known names.
  • mxfp4 — boolean — True if the model ships natively in MXFP4 (e.g. gpt-oss)
  • budget — number — Optional max price in USD
  • context — number — Context window in tokens (default 8192)
  • total_b — number — For an unlisted model: total parameters in billions
  • active_b — number — For an unlisted model: active params in billions (= total for dense, less for MoE)
  • kv_precision — string — KV cache precision (default f16)

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• get_used_gpu_prices

  Current typical used-GPU prices for local-AI rigs (eBay Browse API median asking + hand-verified, monthly).

  Parameters:

  • gpu — string — Optional name/id filter, e.g. "3090"

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• compare_hardware

  Side-by-side memory, bandwidth, price, and (with a model) fit + tok/s for 2 to 4 machines.

  Parameters:

  • hardware (required) — string — 2 to 4 hardware names/ids, comma-separated
  • model — string — Model name, e.g. 'Llama 70B', 'gpt-oss-120B', 'Qwen 32B'. Use list_models to see known names.
  • mxfp4 — boolean — True if the model ships natively in MXFP4 (e.g. gpt-oss)
  • context — number — Context window in tokens (default 8192)
  • total_b — number — For an unlisted model: total parameters in billions
  • active_b — number — For an unlisted model: active params in billions (= total for dense, less for MoE)
  • kv_precision — string — KV cache precision (default f16)

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

Publisher website

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.