The short version: a free, hosted Model Context Protocol server at https://vettedconsumer.com/mcp that lets any MCP client (Claude, Cursor, Cline, Windsurf, or your own agent) ask nine questions about local-LLM hardware: will this model fit this machine, which GGUF quant to download, the cheapest box that runs it, buy versus rent versus API cost, and what used GPUs cost this month. No install, no API key, no account. It runs the exact same engine as the calculators on this site, so an agent gets the same answer a person gets.
Connect in one line
Add this to your client's MCP configuration. Claude Desktop, Cursor, Cline and most other clients accept the same shape:
{
"mcpServers": {
"vetted-consumer": { "url": "https://vettedconsumer.com/mcp" }
}
}Clients that only speak stdio can bridge it with npx mcp-remote https://vettedconsumer.com/mcp. The transport is Streamable HTTP, stateless, with no authentication. A browser visiting the endpoint directly gets a 405, because MCP talks over POST; that is expected, not broken.
The nine tools
| Tool | What it answers |
|---|---|
can_i_run_it | Will a given model run on given hardware? Returns fit, the best quant that fits, a theoretical tokens-per-second ceiling, and an owner-measured figure with its source where one exists. |
recommend_quant | Which GGUF quant to download for a model on your machine: the full ladder with file size, maximum context and speed tier. |
cheapest_hardware_for_model | The cheapest catalogued, buyable machine that runs a model at Q4 with the context you asked for. |
recommend_hardware | A ranked list of buyable machines that run a model at the requested context, cheapest first. |
compare_hardware | Side by side memory, bandwidth and price for two to four machines, plus fit and speed if you name a model. |
cost_compare | Buy versus rent versus API: monthly, one-year and three-year totals, break-even months, and the energy cost per million tokens. |
get_used_gpu_prices | Current typical used-GPU prices for local-AI rigs, from a monthly eBay median plus hand-verified entries. |
list_models | The model classes the tools know: parameters, dense or MoE, active parameters, native context. |
list_hardware | The machines the tools know: memory, bandwidth, price, and whether the memory is unified. |
The same answers over plain HTTP
Every tool is also a GET endpoint under /api/, for scripts and for anyone who wants to check the math without an agent. GET /api lists them. Three real responses, trimmed for length but otherwise unedited:
GET /api/can-i-run-it?model=Llama%2070B&hardware=rtx-3090&context=8192{
"model": {
"name": "Llama 3.x 70B (dense)",
"total_b": 70,
"active_b": 70,
"type": "dense"
},
"hardware": {
"name": "NVIDIA RTX 3090 24GB",
"memory_gb": 24,
"unified_memory": false,
"memory_bandwidth_gbps": 936
},
"fits": false,
"verdict": "Does not fit 8192 context on 24 GB, even at the smallest quant. Use a smaller model or more memory.",
"owner_measured": {
"tok_s_low": 10,
"tok_s_high": 10,
"quant": "Q2_K",
"note": "Llama 70B at Q2_K, the largest quant that fits 24GB (Q4 spills to RAM)",
"sources": 1,
"source_url": "https://mustafa.net/llm-tokens-per-second-benchmarks/"
}
}GET /api/can-i-run-it?model=gpt-oss-20B&hardware=rtx-3090&context=8192{
"model": {
"name": "gpt-oss-20B (MoE, ~3.6B active)",
"total_b": 20,
"active_b": 3.6,
"type": "MoE"
},
"fits": true,
"recommended_quant": {
"quant": "MXFP4",
"file_size_gb": 10.6,
"max_context": 55313,
"tok_s_theoretical": 266,
"speed": "fast"
},
"verdict": "Runs at MXFP4 (~10.6 GB weights), 266 tok/s theoretical (fast)."
}GET /api/cost-compare?hardware=rtx-3090-used&hours=3&tokens=300000{
"hardware": "NVIDIA RTX 3090 24GB (used)",
"price_usd": 1100,
"monthly_cost": {
"buy_electricity": 3.53,
"rent": 53.1,
"api": 9
},
"totals": {
"buy": {
"1y": 1142,
"3y": 1227
},
"rent": {
"1y": 637,
"3y": 1912
},
"api": {
"1y": 108,
"3y": 324
}
},
"cheapest_after": {
"1y": "api",
"3y": "api"
},
"break_even_months": {
"vs_rent": 22.2,
"vs_api": 201
}
}Why the numbers match the website
The server loads the exact specification file and hardware catalogue that the on-site calculators use. An hourly job copies the live theme files into the server and restarts it only if they changed, so the API cannot drift from Can I run it?, the quant picker or the cost calculator. The fit arithmetic is the one shown on those pages: weights at 4.8 bits per weight for Q4_K_M, a KV cache that grows with context, a fixed runtime overhead, and a memory reserve that is larger on unified-memory machines because macOS keeps part of the pool for itself. Speeds are bandwidth ceilings unless an owner has published a measurement, in which case the response carries the measured figure and its source in owner_measured.
What it knows, and what it does not
The catalogue is curated rather than exhaustive: 18 model classes from a 3B dense model to a 671B Mixture-of-Experts, and 23 machines from a used RTX 3060 to a 512GB Mac Studio. A model that is not listed can still be sized by giving its total and active parameter counts. Name matching is forgiving but not clever, so if a result names a different model than you meant, use the exact class name from list_models. We have not benchmarked these machines ourselves; everything not marked measured is a ceiling, and prices are dated to the month they were observed.
Discovery, for the agents that look
- Server card: /.well-known/mcp/server-card.json
- API catalog (RFC 9727): /.well-known/api-catalog, also advertised in an RFC 8288 Link header on every page
- Agent skills index: /.well-known/agent-skills/index.json
- Curated index for language models: /llms.txt; every article is also served as markdown to a client that sends
Accept: text/markdown - DNS: a TXT record at
_agent.vettedconsumer.compoints at the endpoint - Listed in the official MCP Registry as
io.github.TheBaronofAI/vetted-consumer - On Smithery as
hi-10f9/vetted-consumer, where it has been used more than 2,400 times; connection URLhttps://server.smithery.ai/hi-10f9/vetted-consumer
Terms
Free, public and non-commercial, like the rest of this site: no affiliate links, no sponsorship, no payment from any brand we cover. There is no account, no key, and no tracking beyond ordinary server logs. It runs on the same server as the site, with no service-level promise; if the site is up, the endpoint is up. If a number is wrong, tell us and it goes on the corrections log. Our methodology page names the testers and sources the catalogue relies on.