Start MCP Inspector
MCP Inspector runs via npx and requires Node.js 18 or later. Verify the installation withnode --version, then start the inspector:
http://localhost:6274 and opens a browser window automatically.
If the Inspector shows a blank white page after starting, the browser has cached a different version of the UI. Open the Inspector in a private/incognito window or press
Ctrl+Shift+R to force a cache reload.Connect to Gcore MCP Server
MCP Inspector supports two connection methods: stdio transport through the UI configuration, and HTTP transport for debugging scenarios where server logs need monitoring.Stdio transport
The stdio method launches the server directly from the Inspector interface.-
Set Transport Type to
STDIO. - Enter the command:
- Enter the arguments:
- Expand Environment Variables and add:
Leave
GCORE_TOOLS unset. The Inspector then lists the six dynamic tools. A pattern from Tool filtering narrows the operations those tools can reach.
- Click Connect.
McpError: MCP error -32001: Request timed out appears when the Inspector UI cannot reach its proxy process on port 6277. This is a known issue with MCP Inspector v0.21+ on Windows where the proxy process fails to start. Use the HTTP transport method instead. The steps are in the next section.
HTTP transport
HTTP mode allows monitoring server logs in a separate terminal.- Start the server with HTTP transport:
127.0.0.1 and logs that HTTP is read-only unless GCORE_READ_ONLY=false.
- Point the Inspector at the session URL:
- Click Connect.
GCORE_TRANSPORT=sse exits with status 2. GET /sse returns 404. The --http flag is rejected.
Explore available tools
After connecting, the Inspector displays three tabs: Tools (available MCP tools), Resources (MCP resources if configured), and Prompts (MCP prompts if configured). Selecting a tool reveals its description, input schema with required and optional parameters, and a test interface for manual execution. In dynamic mode the Tools tab listssearch_tools, get_tool_schema, call_read_tool, call_write_tool, call_destructive_tool, and execute_code. HTTP transport drops the write and destructive tools and leaves four. GCORE_MCP_MODE=direct with a GCORE_TOOLS pattern lists one tool per operation instead. GCORE_TOOLS=* in direct mode registers 818 tools.
Execute tools manually
Tool execution requires selecting a tool from the list, filling parameter values according to the input schema, and clicking Run Tool. The Results panel displays the JSON response from the server. Selectcall_read_tool, set name to cloud.instances.list, and leave arguments empty. Click Run Tool. The response names the project it used and the region on each instance:
get_tool_schema and set name to cloud.instances.list to read the argument schema, including which fields are optional. call_destructive_tool for dns.zones.delete returns CONFIRMATION REQUIRED and does not send the call.
Understand tool schemas
Tool schemas define the expected input format and appear in the Input Schema section when selecting a tool. Each field includes its data type (string, number, object, or array), whether it is required, a description of its purpose, and default values if applicable. Expanding nested objects reveals all available configuration options. Schema exploration helps understand exact parameter names and formats before constructing requests in Claude Code or other clients.Debug execution errors
The Inspector displays the full JSON-RPC request sent to the server, the response with results or errors, and execution timing. The server log names each API request and its status:Practical applications
MCP Inspector serves as a reference tool when documenting or integrating Gcore MCP Server. Exploring tools reveals exact names, parameter requirements, types, and response structures. Testing filter patterns confirms thatGCORE_TOOLS values match the intended tool set before configuring production clients. When Claude Code reports errors, replicating the tool call in the Inspector reveals the exact request and full error response for debugging.