Sanity Challenge · bring your own agent

What computer do I need to run this AI model?

Ask your own AI agent. Connect ChatGPT, Claude, Codex or Cursor to this server, and it answers from a Sanity knowledge base of formulas, rules and real measurements, then checks its own math against the base before you see the answer. It also tells you when to keep the computer you have or rent instead.

https://hw-advisor.helgardorlm.tech/mcp

A remote MCP server: streamable HTTP, no login. Your model, your subscription; the base and the math check are on us.

One click

Cursor or VS Code
Opens your editor and adds the server.

Add to Cursor Add to VS Code

Any agent that runs commands (Claude Code, Codex, Cursor agent)
Paste this. The agent reads the setup page, offers to install the server for later sessions, and answers now over HTTPS.

Plain chat with web access (ChatGPT free, Claude, Gemini) experimental
No setup. The chat opens our plain HTTPS API by itself. This depends on the chat's browsing tool, which may refuse long query links.

Or connect it by hand

Use ChatGPT on the web with a Plus, Pro, Business, Enterprise or Edu plan. The free plan cannot add custom connectors.
Settings → Apps → Advanced settings: turn on Developer mode.
Click Create. Name: AI Hardware Advisor, URL: the address above, Authentication: None.
In a new chat, pick the connector from the + menu (Developer mode) and ask a question below.

claude mcp add --transport http hw-advisor https://hw-advisor.helgardorlm.tech/mcp

Then run claude and ask a question below.

Add to ~/.codex/config.toml:

[mcp_servers.hw-advisor]
url = "https://hw-advisor.helgardorlm.tech/mcp"

Add to ~/.cursor/mcp.json:

{ "mcpServers": { "hw-advisor": { "url": "https://hw-advisor.helgardorlm.tech/mcp" } } }

Add to .vscode/mcp.json (or run MCP: Add Server → HTTP):

{ "servers": { "hw-advisor": { "type": "http", "url": "https://hw-advisor.helgardorlm.tech/mcp" } } }

Questions to try

Each Copy puts this line in front of the question: "Use the AI Hardware Advisor tools: call initial_context first…". Without it, an agent may answer from memory and skip the base. Agent files: setup.md · SKILL.md (Claude Code and Codex skill) · llms.txt.

What happens when you ask

1Instructions + schemainitial_context gives the agent the advisor's work order and the dataset schema through Sanity Context MCP.

2Need in numbersLaws (weights, KV cache, speed ceiling, own vs rent) are fetched as data with groq_query. Models that are not in the base are looked up on the web, using the base's rules as a checklist.

3Every pathAll 11 solution paths, from keeping what you have to renting a GPU, get a yes, no or maybe with a number.

4Math checkedcheck_answer recomputes each calculation with the formula stored in Sanity and returns errors until the answer is right.

Tools on this server

initial_context	Advisor instructions plus the Sanity schema and document counts. Call it first.
groq_query	Read-only GROQ over the public dataset `v2`, with embeddings on, so semantic search works.
schema_explorer	Fields and allowed values of one document type.
array_field_reader	Reads long array fields in parts.
check_answer	The validator from the app. It recomputes laws, checks that every path got a verdict, catches substitute products and flags prices without a source. Added by this server.

The first four tools are Sanity Context MCP, passed through unchanged. The access token stays on the server.

No paid plan? See it run

A replay of all 27 measured runs shows the full tool calls, the base documents the agent read, and the validator's corrections: helgard-orlm.github.io/sanity-ai-hardware-advisor.

Under the hood

Source on GitHub Sanity Studio (dataset v2) Sanity project onwa0wvs · public dataset v2

Limits and privacy

This is a free demo on a free Sanity plan, so calls are limited per visitor per hour and per day in total. Past the daily limit, groq_query falls back to the public dataset API. The server logs which tool was called and the GROQ query text, with a salted hash in place of your IP address. It never sees your chat. Answers are advice from a knowledge base, not a guarantee. Check prices and specifications before you buy.