Skip to main content

​
Realtime LLM Interaction Logs

Observability is crucial when building LLM agents. Because the model’s decision-making process (“Which tool should I call?”) is probabilistic, you need visibility into exactly what is happening under the hood. HasMCP provides a real-time log stream for every MCP Server you create. This allows you to see:
  • Connection Events: When a client (like Claude) connects or disconnects.
  • Tool Calls: The exact arguments the LLM generated.
  • API Responses: The raw data returned by your external provider.
  • Errors: Authentication failures or schema validation issues.

​
Accessing Logs

  1. Navigate to the MCP Servers tab.
  2. Locate the server you want to monitor.
  3. Click the Logs icon (a document/file icon) in the top-right corner of the server card, or click “View MCP Server Logs” inside the server details page.
Status Indicator: The top of the log window shows the connection status.
  • Connecting…: Establishing the Server-Sent Events (SSE) stream.
  • Streaming…: Live and receiving data.
  • Disconnected: The connection was lost or the tab was closed.

​
Understanding Log Entries

The log viewer displays events in chronological order (newest at the top).

​
1. Connection Lifecycle

When you start your LLM client (e.g., open Claude Desktop), you should see an initialization sequence:
  • connection: init - The client has established an HTTP connection.
  • capabilities: negotiation - The client and server are agreeing on supported features (e.g., resources, tools).

​
2. Request / Response Cycle (The “Thinking” Process)

This is the most critical part for debugging. A typical tool use sequence looks like this: A. The Request (Blue) The LLM decides to call a tool. You will see an event starting with req.
  • Event: req:call_tool
  • Data: {"name": "list_customers", "arguments": {"limit": 5}}
  • What to look for: Did the model hallucinate an argument? Is the data type correct (e.g., string vs integer)?
B. The Response (Green) Your HasMCP server executes the request against the external API and returns the result.
  • Event: res:call_tool
  • Data: {"content": [{"type": "text", "text": "..."}]}
  • What to look for: Did the API return the expected data? Is the JSON too large for the context window?

​
3. Errors (Red)

If something goes wrong, you will see a red error entry.
  • 401 Unauthorized: Usually means your Environment Variable is missing or incorrect.
  • 404 Not Found: The endpoint path might be configured incorrectly in the Provider settings.
  • 500 Internal Error: A parsing error or unexpected failure in the backend.

​
Inspecting Payloads

LLM tool calls and API responses can be large JSON objects. HasMCP truncates these in the main view to keep the log readable. To view the full payload:
  1. Click directly on the Data: {...} text of any log entry.
  2. A modal will open showing the Prettified JSON.
  3. Use the Copy JSON button to paste it into an external editor or validator.

​
Troubleshooting Common Scenarios

​
”I don’t see any logs when I use Claude.”

  • Cause: Claude Desktop might not be running or the configuration file points to the wrong URL.
  • Fix: Restart Claude Desktop. Ensure the mcpServers config in your claude_desktop_config.json matches the snippet provided in the Server Details page.

​
”The model keeps getting ‘Invalid Arguments’ errors.”

  • Cause: The JSON Schema defined in your Provider Endpoint might not match what the API actually expects.
  • Fix: Go to the Providers tab, edit the endpoint, and check the Query Arguments or Request Body schema. Ensure required fields are actually marked as required.

​
”I see a 401 error in the logs, but I added the variable.”

  • Cause: The variable name might not match the Secret Prefix required by the provider.
  • Fix: Go to the Variables page. Check if your variable starts with the exact prefix shown on the Provider Details page (e.g., API_STRIPE_COM_...).