From zero to your
first MCP server.

Ask your technical administrator for these details

1. The database type or a copy of your OpenAPI/Swagger document.
2. A read-only service account where possible.
3. The server address, database name, username, password, and whether encrypted connections are required.
4. The business question your AI should answer—for example, “Find a customer by email.”

What to expect: MCP Forge stores runtime configuration on your infrastructure. Do not use a personal administrator account; use a dedicated account with only the permissions required.

Enter your PostgreSQL connection

1. Open MCP Forge and select Runtime → Queries.
2. Choose Connection, enter a friendly name, and select PostgreSQL.
3. Enter a connection string such as postgresql://report_user:password@db.company.local:5432/sales.
4. Save the connection. Your database administrator may need to allow network access from the MCP Forge computer.

What to expect: PostgreSQL queries use $1, $2 placeholders. Results are limited to 1,000 rows by the runtime.

Enter your MySQL connection

1. Open Runtime → Queries → Connection.
2. Select MySQL and enter a name your team will recognize.
3. Enter mysql://report_user:password@db.company.local:3306/sales.
4. Save, then create approved queries using ? placeholders.

What to expect: MySQL parameters are supplied in the same order as the parameter list you define.

Choose a local database file

1. Open Runtime → Queries → Connection.
2. Select SQLite.
3. Enter the full path to the database file, such as C:\Data\customers.db on Windows or /data/customers.db in Docker.
4. Ensure the MCP Forge process can read the file.

What to expect: SQLite requires no database server. Docker users must mount the database file into the container.

Enter your SQL Server connection

1. Open Runtime → Queries → Connection and select SQL Server.
2. Enter a connection string such as Server=sql.company.local;Database=Sales;User Id=mcp_reader;Password=…;Encrypt=true;TrustServerCertificate=false.
3. Ask your administrator whether SQL authentication is enabled and which encryption settings to use.
4. Save the connection. Define query inputs with SQL Server names such as @email.

What to expect: Each parameter named in MCP Forge is bound to the matching @parameter in your T-SQL query. Use TOP to keep results intentionally small.

Import the API document

1. Open Runtime → API Specs → Import document.
2. Give the API a friendly name.
3. Enter the base URL if the document does not already include one.
4. Paste the complete JSON or YAML document and save.
5. Open Runtime → Endpoint and select Test tools/list to confirm the operations were generated.

What to expect: Every supported operation becomes a tool. Its operationId becomes the tool name; descriptions help the AI decide when to use it.

Create a small OpenAPI wrapper

1. Collect one working request: method, URL, headers, parameters, and example response.
2. Create a minimal OpenAPI document containing only the endpoints you want the AI to call.
3. Use clear operationId and description values.
4. Import the wrapper under Runtime → API Specs.

What to expect: MCP Forge intentionally avoids exposing an entire undocumented API automatically. A small approved wrapper creates a safer, clearer tool set.

Use an approved REST/OpenAPI façade

1. Choose the GraphQL operations you want to make available.
2. Create server-side REST endpoints or an OpenAPI façade that runs those fixed queries and mutations.
3. Describe the inputs and outputs in OpenAPI.
4. Import that document into MCP Forge.

Current support: Direct arbitrary GraphQL documents are not executed by the current runtime. The approved-façade pattern prevents an AI from composing unrestricted GraphQL operations.

1. Open Runtime → Queries → Query.
2. Enter a short action-oriented tool name such as find_customer_by_email.
3. Select the connection and write a plain-language description.
4. Paste the parameterized SQL. Never combine user text directly into SQL.
5. Define a JSON parameter array. Names must match the business inputs and, for SQL Server, the @names in the query.
6. Leave read-only enforcement enabled unless the tool must write data.

[
  {"name":"email","type":"string","required":true,
   "description":"Exact customer email address"}
]

What to expect: The tool appears in tools/list immediately. The AI sees its name, description, and parameter schema—not your database password.

1. Set a long random MCPFORGE_TOKEN before starting MCP Forge.
2. Open Runtime → Endpoint.
3. Use the default endpoint or create a named endpoint such as customer-data.
4. Select Test tools/list. Confirm the expected number of tools.
5. Copy the endpoint URL.

What to expect: A local URL looks like http://localhost:4173/mcp/default. For remote use, place MCP Forge behind HTTPS and an authenticated reverse proxy.

1. Open the MCP or connector settings in your chosen AI client.
2. Create a server named MCP Forge.
3. Paste the endpoint URL.
4. Add Authorization: Bearer YOUR_TOKEN as a request header.
5. Connect and review the tools the client discovers.

{
  "mcpServers": {
    "mcp-forge": {
      "url": "https://mcp.example.com/mcp/default",
      "headers": {"Authorization":"Bearer YOUR_TOKEN"}
    }
  }
}

Check the hostname, port, credentials, encryption settings, firewall, and database permissions. From Docker, localhost refers to the container—not the host computer.

Confirm the query saved successfully or the OpenAPI operation has a supported HTTP method. Then call tools/list again or reconnect the client.

Read-only tools must begin with SELECT, WITH, EXPLAIN, or PRAGMA. For an intentional write, disable read-only only after reviewing permissions and adding an approval workflow.

Email support@builtbykris.com with the connection type and the error message. Never email passwords, tokens, or full connection strings.