MCP Insights

Feature

New in 1.5.0

The MCP Insight view shows specific info about tools,resources,prompts usage, including token count.

Inspectr introduces first-class Model Context Protocol (MCP) insights. MCP requests are auto-detected, categorized, and enriched so you can see which tools, resources, and prompts are being exercised—and how many tokens each call costs.

What you get

• Auto-detection of MCP traffic operations are marked with the mcp protocol.
• Classification by category and type: Inspectr parses method + params to tag calls as tool, resource, prompt, or system (e.g., tools/call, resources/read, prompts/get, initialize, notifications).
• Name extraction: pulls the tool/resource/prompt identifier (search_tool, resource URI, prompt name) into the operation metadata and tags.
• Token accounting: Estimate request + response tokens usage and stores request/response/total per operation.
• MCP-specific metrics: tracks totals and errors for MCP requests, plus counts per tool/resource/prompt.
• Trace-friendly IDs: captures MCP session/request IDs from headers to correlate chat flows and SSE streams.

MCP details per operation

The details panel shows MCP information for each operation:

• Classification info:
  • Tool/Resource/Prompt name (e.g.get_trace)
  • The method (e.g. tools/call)
  • The category (tool, resource, prompt, system) (e.g. tool)
• Token counts (request, response, total) are attached to each MCP operation.
• Input
  • The parameters passed to the MCP method (referenced in the MCP spec as the params field).
• Output
  • The structured response content (referenced in the MCP spec as the structuredContent field)
  • The content response (referenced in the MCP spec as the content field)
  • The raw response body, containing both structured and content items.

Trace View - Usage insights and token estimates

Inspectr aggregates MCP activity per session so you can get insights:

• About the sequence of the MCP interactions
• Which tools are called most often? Which resources are read? Which prompts are popular?
• How many tokens are burned by MCP traffic over a window?
• How often do MCP calls error (HTTP or JSON-RPC)?

Per operation, you can see all the details, like mentioned above.

How classification works

• Methods with tools/* become tool calls; resources/* → resource; prompts/* → prompt; notifications/roots/tasks/logging fall under system.
• Names are pulled from params:
  • tools/call → params.name
  • resources/read → params.uri / first entry in params.uris
  • prompts/get / prompts/call → params.name
• Unknown/other MCP methods still attach protocol and token counts, defaulting category to system.

See MCP in action with LLMs

• Connect with LLMs like ChatGPT, Claude, or any MCP client to connect to your MCP server.
• Watch each MCP call in Request History, including the tool/resource/prompt invoked and its token cost.
• Use the built-in Inspectr MCP server (tools + resources + prompts) to query your own traffic from inside a chat.
• Build your own MCP server using FastMCP and connect to it with Inspectr.

Explore MCP Insights on the product page →