MCPblaze docs
Your first MCP server
How to create an MCP server, change it, delete it, and use the scorecard to get it ready for the Claude and ChatGPT directories. No experience needed. Each step is one command you can copy.
The shortest way: ask your AI assistant
Open Claude Code or Codex in your app’s folder and paste this. It does the steps on this page for you and asks when it needs you.
What you need
An app with an API that is already running, on the internet or on your own computer.
Node.js installed, so that npx works in your terminal.
No account is needed to start. You can sign in later and keep what you made.
What is an MCP server?
Your app has an API: a set of addresses a program can call to read or change things, such as “list my bookings” or “cancel this booking”. An AI assistant like Claude or ChatGPT cannot use an API on its own. It needs an MCP server: a small translator that sits in front of your API and tells the assistant, in a standard way, what it can do for the person chatting.
Tool
One thing the assistant can do with your app, like “book a table”. MCPblaze makes one tool for each call your API offers, and you keep the useful ones.
Sign-in
Each person signs in before the assistant acts for them, so Alice’s assistant only sees Alice’s data. MCPblaze adds this for you.
Directory
The list of apps inside Claude and ChatGPT where people find yours. Each directory has rules. The scorecard checks them.
What MCPblaze does
You give it your API. It gives you a hosted MCP server with sign-in, and a scorecard that says what is left before you can ask Claude and ChatGPT to list it. Your app’s code does not change.
Create one
One command. Pick the line that matches where your app runs.
- 1
Run the create command
Your app is on the internet
Give the address of your API, or of its OpenAPI file if you have one.
npx mcpblaze@latest create --target https://api.example.com/openapi.yaml --name myappYour app runs on this computer
Give the port it listens on. MCPblaze opens a tunnel to it, so keep this command running while you test.
npx mcpblaze@latest dev 3000 --name myappWhat is an OpenAPI file?
A file that lists every call your API offers. Many frameworks write it for you, often at
/openapi.json. With it, MCPblaze knows your tools by name. Without it you still get an MCP server, with fewer details. - 2
Keep the two links it prints
MCP server https://myapp-acme.mcpblaze.com Scorecard https://dashboard.apiblaze.com/directory/myapp/1.0.0
The first is the address an assistant connects to. The second is the page that shows what is left to do. The exact wording in your terminal may differ; the two links are what matter.
- 3
Sign in when you want to keep it
npx mcpblaze@latest loginIt offers to move what you made without an account into your account.
npx mcpblaze@latest projectsLists every MCP server you have.
The name matters
The name becomes part of the address (myapp-yourteam.mcpblaze.com). Keep it short and simple, and do not use a name that belongs to someone else.
Try it in Claude or ChatGPT
See your tools work in a real chat before you change anything.
In Claude
Open Customize, then Connectors, then Add custom connector. Paste your MCP server address, sign in, and ask for something your app can do.
In ChatGPT
Open chatgpt.com/plugins, choose Add custom MCP server, paste your MCP server address, and sign in.
In Claude Code or Codex (the terminal)
Then type /mcp inside Claude Code and sign in.
Then sign in when your browser opens.
Edit it
Changes apply to the server that is already running. You do not create it again.
Why edit at all?
An assistant picks a tool by reading its name and description. A dozen clear tools work better than one tool for every call your API has. Most edits are: remove what nobody would ask for, and describe the rest in plain words.
Choose which tools to keep
Asks about each tool: keep it or remove it.
Removes one tool. Your API is untouched; the assistant just no longer sees it.
Brings it back.
Rename or describe a tool
Shows every tool and how it behaves.
A new name and a description the assistant can understand.
Say how risky a tool is
Assistants treat tools differently when they know what a tool does. MCPblaze guesses from your API (reading is safe, deleting is not). Correct it when the guess is wrong.
Only reads, changes nothing.
Deletes or changes data, so the assistant asks the person first.
Back to the automatic setting.
Build one tool around what people ask for
Sometimes the thing people want takes several API calls. This makes one tool that does them in order.
It proposes a goal; you accept it or write your own.
Tell the assistant what your server is for
A short note every assistant reads when it connects.
Shows your whole server the way an assistant reads it. Good for a last check.
Change who can sign in
Opens a picker: GitHub, Google, or your own sign-in.
Your API changed?
After a big change to your API, run mcp ops list myapp to see that every tool still looks right, and check the scorecard again.
Use the scorecard
The scorecard is a to-do list with a score. It tells you what the Claude and ChatGPT directories check, and what is still missing.
Open it
Use the Scorecard link from the create step. It looks like this, with your own name in it:
https://dashboard.apiblaze.com/directory/myapp/1.0.0
Lost the link? This prints the score and the link again:
The score
A number out of 100. It goes up as steps are completed. 100 means every check we can run has passed.
The steps
Twelve steps, in order. Each is done, open, or waiting for an earlier step.
The fix
Every open step says what is missing and how to fix it, on the page or as a command to copy.
The twelve steps:
- 1
Get listed
Say which directories you want: Claude, ChatGPT, or both.
- 2
Your hosted LLM Chat server (MCP)
The address of your MCP server, and the hosted app it talks to.
- 3
How people sign in to your MCP
Sign-in for each person, or open to everyone for public read-only data.
- 4
What your App can do in chat
Your tools: clear names, clear descriptions, the right safety labels.
- 5
Your listing page
Name, description, logo, example prompts, privacy and support links.
- 6
Listing questionnaire
A few questions the directories ask about your app and its data.
- 7
What you confirm
Statements you tick to say you follow the directory policies.
- 8
Test your MCP
We call every tool and try your example prompts, and keep the result as proof.
- 9
Connect it to Claude yourself
You add it to your own Claude once, sign in, and use one tool.
- 10
Prove you own this domain (ChatGPT)
Only for ChatGPT: a short code that shows the domain is yours.
- 11
Create your App's submission
One packet with everything the directory form asks for.
- 12
Send it to Claude
You paste the packet into the directory form and send it.
Do the steps from the terminal instead
This walks through every open step in order and asks for what is missing. Stop any time; it picks up where you left off.
You need your own web address
The directories want the MCP server on a domain you own, such as mcp.yourdomain.com, and an app that is on the internet, not on your computer. Step 2 walks you through both. A custom domain is billed monthly; see Pricing.
Running the tests
In step 8 we call each tool and try your example prompts with a real AI model. The first 60 model rounds for a listing are free. See Pricing for what comes after.
You submit. They decide.
A full score means your MCP server meets the requirements we can check. It is not a listing. You send the submission yourself, and Anthropic and OpenAI decide whether to list your app.
Delete it
Deleting removes the MCP server and everything under it.
- 1
Find its name
npx mcpblaze@latest projects - 2
Delete it
npx mcpblaze@latest delete myappIt asks you to confirm first.
This cannot be undone
The address stops working at once, so anyone who connected it in Claude or ChatGPT loses it. Its tools, sign-in settings, keys and scorecard progress go with it. Your own app and its API are not touched.
Only want to stop one tool?
You do not need to delete the server. Hide the tool instead: mcp ops hide myapp <tool>.
Words used on this page
- API
- The addresses a program calls to use your app.
- MCP
- Model Context Protocol: the standard way an AI assistant talks to an app.
- MCP server
- The translator between an assistant and your API. MCPblaze hosts it for you.
- Tool
- One action the assistant can take through your MCP server.
- Connector
- What Claude calls an MCP server once someone has added it.
- Directory
- The list of apps inside Claude or ChatGPT.
- Scorecard
- The page that shows how ready your MCP server is for a directory.
- Terminal
- The window where you type commands. On a Mac it is the Terminal app.
- npx
- Runs a tool without installing it. It comes with Node.js.
Stuck?
Ask in your terminal
A chatbot that knows MCPblaze and can act on your account.
Every command, with one line each.
Ask us
Write to support@apiblaze.com.
MCPblaze is in beta: things can change or break. See the Terms.
For API keys, rate limits and everything else the gateway does, read the APIblaze docs.