thinkhard.si prompt tools

npx add-mcp 'https://thinkhard.si/mcp'

Add it to your assistant

New to connectors? MCP servers for people who don’t code explains them in plain words, with what to check before you add one.

1. In Claude on the web or the desktop app, open Customize, then Connectors, and choose Add, then Add custom connector.
2. Name it thinkhard.si, paste https://thinkhard.si/mcp, and under Authentication choose No sign in.
3. In a chat, turn it on from the + menu under Connectors and ask, for example, “Find a thinkhard prompt for a cover letter and fill it in with my resume.”

Works on every plan, including Free, which allows one custom connector. On Team and Enterprise, an owner adds it once under Organization settings, then Connectors, and everyone can turn it on. A connector added once works in Claude on the web, the desktop app and the phone.

Prefer the version that runs on your computer, with the two tools that read your drafts? Download the Claude Desktop extension and open it. Claude Desktop installs it, with no Node or terminal needed.

Try it from a terminal

No setup, no session: one request grades a prompt and returns the score and what is missing.

curl -s https://thinkhard.si/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"grade_prompt","arguments":{"prompt":"Summarize these meeting notes."}}}'

Not using MCP? The same tools are a plain REST API, with an OpenAPI description:

curl -s 'https://thinkhard.si/api/v1/prompts?q=meeting+minutes&limit=3'

Or run any tool from the playground in your browser, and copy the same call as curl, JavaScript or Python.

How a call works

What happens when your assistant uses thinkhard.si

1. In your AI app. You ask “Find a prompt for my performance self-review.”
2. In your AI app. Your assistant picks a tool search_prompts, with your words as the search
3. At thinkhard.si/mcp. One request reaches the server A single POST: no session, no key, nothing stored about 100 ms
4. At thinkhard.si/mcp. The answer comes back Text for the model to read, and the same answer as data
5. In your AI app. Your assistant replies The prompt, filled in from your conversation

Your assistant decides when a tool would help, the way a colleague decides when to look something up. Everything else about the conversation stays with your AI app; thinkhard.si only sees the call itself, and keeps a daily count of calls per tool.

What your assistant can do with it

Tool	What it does	Runs
search_prompts	Finds ready-made prompts in the repository for a job, such as “meeting minutes” or “cancel a subscription”.	Both
get_prompt	Fetches one prompt with its blanks and why it works, and fills in the blanks you give it.	Both
list_collections	Lists the repository’s collections and how many prompts each has.	Both
grade_prompt	Scores a prompt out of 10 on the ten checks and asks a plain question for each one it misses, exactly as the Grader does.	Both
estimate_ai_cost	Prices a workload on every current model from the price index, per call and per month.	Both
check_robot_speak	Finds what makes a draft sound like AI, fixes the safe ones and writes instructions that stop them next time, like the Robot-Speak Checker.	Local
fact_check_list	Lists every number, quote, source and link in an AI answer, riskiest first, like the Fact Check List.	Local

Every prompt in the repository is also offered as an MCP prompt, with its blanks as the questions, so clients that support prompts show them as slash commands or in a menu. The tools reference has every parameter, the fields each tool returns, and real answers.

Things to ask

• “Find a thinkhard prompt for a performance self-assessment and fill it in from my notes.”
• “Grade the prompt I just wrote with thinkhard, then fix what it says is missing.”
• “What would 20,000 support replies a month cost on the cheapest models? Use thinkhard’s prices.”
• “Before I send this, check it for robot-speak.” (local server)

The local server

The local server runs on your computer with Node 18 or later and has no dependencies. It runs the website’s own code for grade_prompt, check_robot_speak and fact_check_list, so the drafts you check never leave your machine, and it asks thinkhard.si only for the repository and prices.

Claude Desktop: download the extension (about 60 KB) and open it, or drag it onto Claude Desktop. It brings its own setup, so there is no Node or terminal involved, and you can switch it off under Settings, then Extensions.

Claude Code:

claude mcp add thinkhard -- npx -y --package=https://thinkhard.si/downloads/thinkhard-mcp.tgz thinkhard-mcp

Cursor, Claude Desktop’s config file and other clients that start a command:

{
  "mcpServers": {
    "thinkhard": {
      "command": "npx",
      "args": ["-y", "--package=https://thinkhard.si/downloads/thinkhard-mcp.tgz", "thinkhard-mcp"]
    }
  }
}

Download the package (version 1.1.0, about 30 KB) to read it first, or point the configuration at the saved file. If an app cannot find npx, give the full path that which npx prints.

Build it into your own app

The server is plain JSON over HTTPS with CORS open, so it works from a script, a backend or a web page. Use it from code has working examples for:

• the REST API: five plain JSON endpoints with an OpenAPI description, for scripts, no-code tools and GPT actions;
• curl, JavaScript and Python straight to the MCP server, with no library at all;
• the official MCP SDKs for TypeScript and Python;
• the Claude API and the OpenAI Responses API, where the model calls the tools itself.

For coding assistants, the whole reference is one Markdown file: /developers/mcp.md.

If it does not show up

• Cursor shows a red dot beside the server: the JSON or the address has a typo. Compare it with the snippet above, save, and start a new chat, since tools load when a chat starts.
• Claude has no Add custom connector option: on Team and Enterprise, only an owner can add one, under Organization settings. On Free, the one custom connector you are allowed may already be in use.
• ChatGPT asks you to sign in: set Authentication to No authentication when you create it. If you cannot create one at all, your plan or your workspace admin does not allow custom MCP servers yet.
• The tools are there but never used: name the server in your message, as in “use thinkhard to…”. Assistants reach for a tool more readily when you mention it.
• The local server will not start: the app cannot find Node. Put the full path from which npx in command, or use the Claude Desktop extension, which needs no Node.
• Still stuck? Check the server answers with the curl command above. If it does not, write to contact@thinkhard.si.

What reaches thinkhard.si

• The server address. What your assistant sends with each call, such as your search words or the prompt to grade, is used to answer that call and is never stored or logged. We keep only a daily count of calls per tool. Your assistant’s own app sees the conversation as it always does.
• The local server. Drafts and prompts you check stay on your computer. Only searches, prompt names and token counts for the repository and prices go to thinkhard.si.
• The prompts. Free to use, change and run in your own work and your clients’ work. Please do not republish the collection or sell it (terms).

More in the privacy policy. For AI assistants that read websites, /llms.txt lists every page in plain text.

Questions

Is the thinkhard.si MCP server free?

Yes. There is no account, no key and no charge. One network can make 600 calls an hour, which is far more than an assistant makes in a working day.

Does it need an API key or OAuth?

No. Add the address and it works. The server never calls an AI model itself: its tools read the repository and the price catalog, or check a prompt with fixed rules, so there is nothing to pay for and nothing to sign in to.

Which apps does it work with?

Any MCP client that can reach a remote server over HTTP, including Claude, ChatGPT, Claude Code, Cursor, VS Code, Codex, Gemini CLI, Devin Desktop (formerly Windsurf), Zed and Cline. For a client that only starts local commands, use the local server, which runs with npx, or the Claude Desktop extension.

What does the server keep?

A daily count of calls per tool, and nothing else. The arguments of each call, such as a prompt to grade, are used to answer it and are never stored or logged.

Why do two tools run only in the local server?

check_robot_speak and fact_check_list read your drafts and AI answers in full. In the local server they run on your computer, so that text never leaves it.

Is there a rate limit?

600 calls an hour from one network. Past that the server answers HTTP 429 and says when to try again. An assistant in normal use makes a few dozen calls a day.

Can I use it from the Claude API or the OpenAI API?

Yes. Both APIs can call a remote MCP server for you, as shown in Use it from code. They use MCP tools only, not MCP prompts, so the templates come through search_prompts and get_prompt.

Can I use it in my own product?

Yes. Call it from your code or give it to Claude or GPT through their APIs, as shown in Use it from code. The prompts are free to use in your work and your clients’ work; please do not republish the collection or sell it.