MCP server for AI assistants

Written by Jakub Rusinowski · Last updated September 30, 2026

Ask your AI assistant "Can my RTX 4090 run Qwen3 32B?" and have it answer from this site's own calculations instead of from memory. The server is free, needs no sign-in, and only reads: it cannot change anything, buy anything or run anything.

The server address

https://elgadmisijrfowuujcmu.supabase.co/functions/v1/mcp

Streamable HTTP, no authentication.

What it does

Read-onlyAll five tools only look things up and compute. None can write, delete, send or install anything, and none reaches the internet to answer.

The same maths as the siteFit, memory and speed come from the same calculation module the rest of the site uses, not from a second copy. Memory is sized from all of a model's parameters and speed from the ones it reads per token, so a Mixture-of-Experts model is not mistaken for a small one.

Unknown stays unknownIf this catalogue does not hold a value, such as a price, the answer says null. The server never fills a gap from general knowledge.

Labelled, never bareSpeeds are ranges with a label: estimated (modelled), community (measured by readers of this site) or measured. See the methodology for how each is produced.

A snapshot, with its dateAnswers come from a copy of the site's data taken when the server was last deployed, so it can trail the site by a few days. Every answer states the date of its data.

Connect it

Claude Code

Run this once in a terminal:

claude mcp add --transport http llmconfigurator https://elgadmisijrfowuujcmu.supabase.co/functions/v1/mcp

ChatGPT (developer mode)

Following OpenAI's documentation for testing an MCP server. Developer mode is not available on every account or workspace.

Open Settings, then Security and login, and turn on Developer mode.
Go to chatgpt.com/plugins and select the plus button.
Enter a name and a short description.
Under Connection, enter the server URL including the /mcp path: https://elgadmisijrfowuujcmu.supabase.co/functions/v1/mcp
Create the connection, then start a new conversation and add the connection from the tools menu.

Other clients and Codex-style config

Any client that supports remote MCP servers over Streamable HTTP can use the URL. Where it takes a JSON config, this is the common shape:

{
  "mcpServers": {
    "llmconfigurator": {
      "url": "https://elgadmisijrfowuujcmu.supabase.co/functions/v1/mcp"
    }
  }
}

No API key, token or header is needed. The exact file and field names differ by client; check its documentation.

Try it without an assistant

The official MCP Inspector can call every tool directly:

npx @modelcontextprotocol/inspector --cli https://elgadmisijrfowuujcmu.supabase.co/functions/v1/mcp --transport http --method tools/list

The five tools

search_catalog — Find hardware or a modelUse this first when the user names a GPU, Mac, mini PC or language model and you need its id for the other tools. Matches names and common aliases in a fixed catalogue of local-LLM hardware and models. It does not return specs, fit or speed, and it does not search the web.

check_hardware_fit — Check whether a model fits hardwareUse this when the user asks whether one specific model runs on one specific machine, at what memory cost, and roughly how fast. Needs a hardware id and a model id from search_catalog. Returns a fit verdict, the memory arithmetic, and a decode-speed range with a confidence label. Memory is sized from all parameters, speed from active parameters. It does not recommend purchases and does not run anything.

models_for_hardware — List models a machine can runUse this when the user asks what models a given machine can run. Needs a hardware id from search_catalog. Returns up to 20 catalogue models ranked for that machine, each with a fit verdict and a speed range, optionally filtered to one use such as coding. It covers only models in the catalogue, at one quantisation and context length per call, and leaves out models that do not fit.

compare_hardware — Compare two to four machinesUse this when the user compares two to four machines. Needs hardware ids from search_catalog. Returns memory, bandwidth, memory type, release year, status and a dated price where one is recorded, side by side. A price of null means none is on record. It does not check any model against the machines; use check_hardware_fit for that.

get_model_specs — Get a model's specificationsUse this when the user asks about one model's size, architecture, context window or licence. Needs a model id from search_catalog. Returns total and active parameters, weight size per quantisation, context length, licence, release date and sources. It does not say whether the model fits any machine; use check_hardware_fit for that.

Things to ask

“Can an RTX 4090 run Qwen3 32B?”
It fits, tightly, with the memory arithmetic and a speed range labelled as an estimate.
“I have a Mac Studio with 64 GB. Which models can I run?”
A ranked list for that machine, each with a fit verdict and a speed range.
“Will Llama 4 Scout run on a 24 GB card?”
No: it needs about 65.8 GB because every expert must be resident, even though only 17B parameters are active per token.
“Compare the RTX 4090 and the RTX 5090 for local LLMs.”
Memory, bandwidth and any price on record with its date. A price this catalogue does not hold comes back as null.
“What licence does Qwen3 30B-A3B use, and how large is it at Q4?”
Total and active parameters, the weight size at each quantisation, context length and licence.

Privacy

No account and no sign-in. To prevent abuse and to count usage, the server keeps an anonymous record of each request:

the day (UTC), with no time of day
the kind of request and which of the five tools was called
a rough client type (for example chatgpt, claude, codex, cursor or other), taken from a keyword in the user agent, which is then discarded
whether it succeeded, and how long it took in milliseconds
a 16-character hash of the network address, built with a secret key that changes its output every day, so the same person cannot be followed from one day to the next

The server does not record:

what you asked: tool arguments and search text are never stored or logged
your network address, your user-agent string, a request or session identifier, or any cookie of its own

These records are deleted after 30 days.

The service that hosts the endpoint (Supabase, behind Cloudflare) keeps its own request logs, which include network addresses and similar technical details. Those are outside this site's control and follow that provider's policies.

Methodology: how every number is calculated | Check your hardware | Privacy policy