Skip to main content

Connect AI agents to WhaTap via MCP

WhaTap accumulates rich monitoring data, but it's only accessible to people who know "which menu to look at, and how." MCP (Model Context Protocol) dissolves this barrier. When an AI assistant like Claude, ChatGPT, or Gemini connects to WhaTap via MCP, users can query server CPU, APM TPS, K8s Pod status, or slow queries in a single line of natural language.

Three things MCP changes​

① Lower the entry barrier​

Old workflow:

"Which menu do I go through to see K8s Pod restart history?"

MCP workflow:

"Show me the Pods that restarted in the last hour."

You can access data without knowing the product's menu structure. New hires, cross-team collaborators, or executives can get answers simply by asking.

② Bring monitoring into AI workflows​

Existing AI agents (code writing, document cleanup, issue triage) can now include real-time monitoring data in the loop.

  • "Did error rate change after this PR deploy?" → AI estimates deploy time and writes the error-rate comparison report
  • "Draft last week's incident postmortem" → auto-collects event history and related traces
  • "Pick the event rules with the noisiest Slack alerts" → frequency analysis + tuning proposals

③ Cross-product queries in natural language​

Connect multiple domains in one question without switching dashboards:

  • "Any API that slowed down during the last 30-minute CPU spike?" — Server × APM
  • "Correlate K8s events with the DB connection pool exhaustion timeline" — K8s × DB

Supported clients​

Table | Supported AI clients
AI clientVendorSetup
Claude DesktopAnthropicJSON config file
Claude CodeAnthropicCLI one-liner
Codex CLIOpenAICLI or TOML
Gemini CLIGoogleCLI or JSON

Setup details: MCP client setup

Adoption steps​

Step 1. Install and connect​

  1. Configure the WhaTap MCP server connection — Getting started with MCP
  2. Register in the chosen AI client — Client setup
  3. Set project code (pcode) and API token environment variables — Environment variables

Step 2. First question​

The canonical first check is "Show me my project list." If the project code appears, the connection is fine.

Common follow-up question patterns:

Table | Common natural-language question patterns
SituationExample question
Server summary"Summarize server CPU and memory for the last 5 minutes"
APM anomaly"Find agents showing anomalies"
Error status"Summarize errors in the last 5 minutes"
K8s check"Check Pod status"
Service relations"Show the service topology"
Auto-generate PromQL"Write a PromQL query for per-Pod CPU usage"
Agent install help"How do I install the APM agent in project 12345?"

More examples: Usage examples

Step 3. Weave into team workflows​

Once connected, recurring work can be unified in natural language.

  • Weekly report draft — "Summarize last week's APM key metrics and fill in the weekly report template" → see the MCP section of the Performance reporting scenario
  • Incident response assist — after getting an alert, "Summarize the transaction traces related to this event"
  • Self-diagnosis for developers — after deploying, "Any metric change for the service I just deployed?"

Advanced patterns​

Integrate with agent pipelines​

Once MCP is in place, dev agents like Claude Code or Cursor can use WhaTap data as decision input.

  • PR review agent: "Check recent performance trends for services in this PR's scope, then comment"
  • Release verification agent: "Auto-compare metrics 1 hour after deploy → create an issue if regression"

Spread within the organization​

  • Non-engineering roles (PM, QA, execs) can access data in natural language → removes the engineer bottleneck
  • Onboarding new hires — before learning the menu structure, let them ask "What does our service look like?"

Custom prompt templates​

Save repeat questions as stored prompts and share in the team wiki:

"Summarize incident events for {project} in the last {period}, sort by recurrence count, and add 3 prevention suggestions."

Teammates can reuse the same template verbatim — raising the floor of analysis quality.

Security and scope​

  • The MCP server authenticates via API token (not shared login credentials)
  • Data access is limited to what the token permits → you can restrict to read-only tokens
  • AI clients talk to the MCP server locally (the conversation itself follows the AI provider's policy)

Troubleshooting: MCP troubleshooting

Current version & open source​

  • Version: v1.2.1
  • Open source: github.com/whatap/whatap-open-mcp
  • Supported tools: 10 (project management 3, data exploration 3, composite analysis 2, install/PromQL 2)
  • Supported domains: Server, APM, Kubernetes, Database, Log

Next steps​

Last updated on