You have a notes app full of prompts you never reuse, a bookmark folder you never open, and a chat history you scroll through hoping to find that one good version. prompts.chat fixes this: it is a free, open-source prompt library with an MCP server, which means your coding agent can search it and pull the right prompt on its own while you work.
One quick correction before we start. The video this post builds on calls the library Promptbase. The project in question is prompts.chat, formerly known as Awesome ChatGPT Prompts. PromptBase is a different thing, a marketplace where people buy and sell prompts. Searching for the wrong name will land you in the wrong place.
Here is what it is, how to connect it to your editor in about five minutes, and where it falls short.
What prompts.chat Actually Is
Fatih Kadir Akin started it in December 2022, right as ChatGPT was catching on, as a simple list of role-play prompts. It has since grown into a full community platform with more than 170,000 GitHub stars. The README describes it as the world’s largest open-source prompt library, and the licensing backs up the free part: the code is MIT, and the prompt content itself is dedicated to the public domain under CC0.
The website is organized into prompts, skills, workflows and categories. Prompts cover text, images, video and audio. Skills are multi-file packages, the kind a coding agent loads from a SKILL.md file. Anyone can submit, vote and browse without paying.
One thing worth knowing up front: the video calls these tested prompts. They are community submitted and community voted. Popular does not always mean good, so treat the library as a very large starting point, not a quality guarantee.
Why the MCP Connection Is the Real Upgrade
A prompt library on a website still leaves you doing the work. You search, copy, paste, fill in the blanks, and repeat tomorrow.
MCP, the Model Context Protocol, changes who does the searching. The prompts.chat server exposes the whole library to any MCP-compatible client. According to the API documentation, your agent gets these tools:
- search_prompts: finds prompts by keyword, with filters for type, category and tag
- get_prompt: fetches one prompt by ID and fills in any variables it contains
- get_skill: pulls a full agent skill, including its files
- save_prompt, save_skill and improve_prompt: available once you add a free API key
All public prompts are also exposed as native MCP prompts, so clients that support slash commands or prompt pickers can list them directly.
What this really means: instead of you hunting for a code review prompt, you tell your agent to find one and run it on the file you have open. The search happens inside the conversation.
How to Connect It (Step by Step)
No account or key is needed for searching and reading. The remote server lives at one address:
https://prompts.chat/api/mcp
Pick your editor below. You only need one.
Option 1: Claude Code
- Open your terminal in any project folder.
- Run this command:
claude mcp add --transport http --scope user prompts-chat https://prompts.chat/api/mcpThe user scope makes the server available in every project, not just the current folder. - Confirm it registered:
claude mcp list - Start Claude Code and type
/mcpto see the server and its connection status.
The general method is covered in the Claude Code MCP documentation. Our Claude Code archive has more setup guides if you are new to the tool.
Option 2: Cursor
- Open or create
~/.cursor/mcp.jsonfor all projects, or.cursor/mcp.jsoninside one project. - Add this block:
{ "mcpServers": { "prompts.chat": { "url": "https://prompts.chat/api/mcp" } } } - Save the file and check Cursor’s MCP settings to confirm the server shows as connected.
Option 3: VS Code
- Create
.vscode/mcp.jsonin your workspace, or run MCP: Add Server from the Command Palette and choose HTTP. - Paste this:
{ "servers": { "prompts.chat": { "type": "http", "url": "https://prompts.chat/api/mcp" } } } - Start the server from the inline prompt that appears above the entry, then open Copilot’s agent mode.
Option 4: Windsurf
- Open
~/.codeium/windsurf/mcp_config.json. - Add this block. Notice the key is
serverUrl, noturl. That one word trips up a lot of people:{ "mcpServers": { "prompts.chat": { "serverUrl": "https://prompts.chat/api/mcp" } } } - Refresh the MCP list in Cascade’s settings.
Prefer to run it on your own machine? The project also ships a local version through npx -y prompts.chat mcp, and a Claude Code plugin you can add with /plugin marketplace add f/prompts.chat.
Test It in 60 Seconds
Do not trust that it works. Prove it. This takes one minute, and it saves you from wondering later why the library never seems to get used.
- Open a project with any code file in it.
- Ask by name: Use the prompts.chat search_prompts tool to find a code review prompt, then apply it to this file.
- Watch for the tool call. Your editor should show a search_prompts request and then a get_prompt request.
- Check the result. If the agent used a prompt from the library, the connection works.
Say the tool name in your first request. Agents choose tools on their own judgment, and until you see it fire once, you do not know whether it will reach for the library unprompted. After that first success, shorter requests like find me a prompt for this usually work, though it is worth checking the tool call now and then.
Three Requests Worth Copying
These work because they tell the agent two things: what to search for and what to do with the result. Vague requests like help me write better get vague results, with or without a library behind them.
- For code: Search prompts.chat for a code review prompt, fetch the best match, and run it on the files I changed today.
- For writing: Find a prompt on prompts.chat for editing a blog post for clarity, show me the prompt first, then apply it to my draft.
- For research: Search prompts.chat for a fact-checking prompt and use it to list every claim in this document that needs a source.
The second request asks to see the prompt first. Do that often. Reading what the agent is about to run takes ten seconds and teaches you more about prompting than any course will.
Honest Limitations
This setup is useful, but it has real edges.
- Quality varies. Anyone can submit. You will find excellent prompts next to mediocre ones, and votes are only a rough signal.
- Library prompts are outside text entering your agent. If your agent can edit files or run commands, skim a prompt before it executes. Instructions hidden in third-party content are the core idea behind prompt injection, which OWASP ranks as the top risk for LLM applications. That risk is low here, but it is not zero.
- Setup differs by editor. The config key and file location change from tool to tool, and these tools update often. If a config stops working, check your editor’s current MCP documentation.
- Some tools need a key. Saving prompts and improving prompts require an API key from your prompts.chat settings. The improve tool runs on their side, so keep confidential text out of it.
- It is a shared public service. The docs ask for respectful request volumes and point heavy users to self-hosting.
- The agent decides when to search. It will not always reach for the library unless you ask.
Quick Reference
| Tool | Where the config goes | Key detail |
|---|---|---|
| Claude Code | claude mcp add command | Use --transport http |
| Cursor | ~/.cursor/mcp.json | Uses url |
| VS Code | .vscode/mcp.json | Top-level key is servers |
| Windsurf | ~/.codeium/windsurf/mcp_config.json | Uses serverUrl |
Is It Worth Five Minutes?
Yes, if you use an AI coding agent and repeat any kind of task, which is nearly everyone. The install costs nothing, takes one command or one small file, and removes the copy-and-paste loop. Treat what it returns as a draft to read, not a command to trust, and you get most of the benefit with little of the risk.
The video’s creator offers a full walkthrough to anyone who comments Prompts. You do not need to wait for it. The steps above cover the same ground, and you can finish before the comment gets a reply.
Know someone who keeps rewriting the same prompt every week? Send them this post, then tell me which prompt your agent found first.
Leave a Reply