Documentation

Generate a working MCP (Model Context Protocol) server that connects an AI assistant to your files, database or API, with access rules you describe in plain language.

How a build works

1. The AI model (open weights, served by Featherless.ai) reads your instructions and turns them into a small JSON policy: limits, blocked fields and tool descriptions.
2. The policy is validated and clamped by the backend. Secret-looking names such as password or token are always blocked.
3. The code comes from tested templates, not from the model, so the project compiles and the security checks are always present.
4. You get a spoken report (optional) and download the .zip from the result page.

You choose a source, list what may be exposed and write your instructions. The service returns a .zip with a TypeScript server, its configuration and a README.

Install the generated server

Requires Node.js 18 or newer. Unzip the download, then:

1. Install and build

npm install
npm run build

2. Connect it to Claude Desktop

Add the server to claude_desktop_config.json. Use the real absolute path of the folder and set the environment variable for your source (see the table in Configuration).

{
  "mcpServers": {
    "clinica-pacientes": {
      "command": "node",
      "args": ["/absolute/path/to/clinica-pacientes/dist/index.js"],
      "env": {
        "DATABASE_URL": "postgresql://readonly_user:password@localhost/clinic"
      }
    }
  }
}

3. Restart Claude Desktop

Quit Claude Desktop completely and open it again. The tools of your server appear in the tools list. Cursor, Windsurf and any other MCP client can run the same command.

Configuration

Every generated project has an mcp.config.json. It holds the resources you exposed and the policy the server enforces. Edit it, then run npm run build again.

Example for a database source

{
  "server": { "name": "clinica-pacientes", "version": "1.0.0" },
  "resources": ["patients", "appointments"],
  "db_engine": "postgresql",
  "policy": {
    "max_rows": 100,
    "blocked_columns": ["password", "token", "diagnosis"],
    "statement_timeout_ms": 15000
  },
  "tool_descriptions": {
    "run_query": "Read-only SELECT over patients and appointments."
  }
}

Policy fields

Environment variables

Secrets never go in mcp.config.json. The server reads them from its environment, which is the env block in the client configuration or a local .env file.

Local Files

Lets the assistant list, read and search text files inside folders you choose. It cannot leave those folders.

• list_files: lists the allowed files.
• read_file: returns the content of one file, up to the size limit.
• search_files: finds text across the allowed files.

Databases

PostgreSQL or MySQL/MariaDB. The assistant can inspect the tables you exposed and run read-only queries.

• list_tables: lists the exposed tables.
• describe_table: shows the visible columns of a table.
• run_query: runs one SELECT statement.

External APIs

Lets the assistant call the endpoints you list on one REST API.

• list_endpoints: lists the allowed endpoints.
• call_endpoint: calls one allowed endpoint and returns the JSON, truncated if long.

Accessibility

MCP Builder is designed so that a blind or low-vision developer can build a server without looking at the screen. This section explains how.

Spoken build report

When a build finishes, the service creates an audio report with an ElevenLabs voice in English, German or Spanish. It says the server name, how many files were created and which rules the server enforces. Choose the language in the voice language field.

Text version of every report

The same report is always written as text in a live region. A screen reader announces it even if audio is off or fails, and a braille display shows it.

Keyboard and labels

Every control can be reached with Tab. Dialogs move focus inside and return it when they close, Escape closes them, every field has a label, and status changes are announced. Automated checks with axe-core report zero violations on all pages in the three languages.

Automatic reading

The result page can play the report by itself. The option to read it aloud automatically can be turned off, for example if your screen reader already speaks every change.

Step by step with a screen reader

1. Open the home page and jump by headings to the build section, for example with the H key in NVDA or JAWS.
2. Fill the fields in order: server name, source type, what to expose, instructions and voice language. Each field has a label and a hint.
3. Activate the build button. Progress and errors are announced, so there is no need to look for them.
4. On the result page the report is announced and can be played. Then activate the download button for the .zip.
5. Open README.md inside the .zip and follow it. It is plain text, with the exact commands.

Troubleshooting

Frequently asked questions