scouter-mcp

한국어 문서는 README.ko.md 를 참고하세요.

A stdio MCP server that connects directly to a Scouter Collector over TCP and queries XLogs, counters, and objects. Its purpose is to let an AI quickly explore Scouter metrics and diagnose root causes. Each result carries txid / gxid / objName / endTimeIso, which you can use as keys to cross-analyze with other observability tools such as OpenSearch or Datadog.

Architecture

Java 17. It reuses scouter-common and ports the scouter.webapp net/server classes into the scouter.mcp.client package. MCP uses the Java SDK 2.0.0 stdio transport. All operations against the Collector are read-only.

Build

./gradlew shadowJar
# output: build/libs/scouter-mcp-<version>-all.jar

The .mcpb bundle is produced only by the release CI (which wraps this jar); local builds just produce the jar.

Registration (Claude Code, Claude Desktop, ...)

For a single collector, install the .mcpb bundle from the GitHub Release for a one-click setup — or copy .mcp.json.example, point it at the fat jar (downloaded from the release or built locally), and fill in the credentials. For multiple collectors, see Multiple collectors.

Multiple collectors

The official release ships one .mcpb bundle (one-click install, a single collector) plus the standalone fat jar. A .mcpb defines exactly one server with one credential set, so it cannot register two collectors at once. For multiple collectors — which usually differ in their whole connection set (host/port and user/password) — use the jar directly and add one entry per collector:

1. Download scouter-mcp-<version>-all.jar from the GitHub Release.
2. Add one mcpServers entry per collector to your client config (.mcp.json / claude_desktop_config.json), all pointing at that same jar, each with its own env set. This keeps each credential isolated:

{
  "mcpServers": {
    "scouter-prod": {
      "command": "java",
      "args": ["-jar", "/ABSOLUTE/PATH/scouter-mcp-<version>-all.jar"],
      "env": {
        "SCOUTER_COLLECTOR_HOST": "prod-collector", "SCOUTER_COLLECTOR_PORT": "6100",
        "SCOUTER_USER": "prod-user", "SCOUTER_PASSWORD": "***", "SCOUTER_TZ": "Asia/Seoul"
      }
    },
    "scouter-stg": {
      "command": "java",
      "args": ["-jar", "/ABSOLUTE/PATH/scouter-mcp-<version>-all.jar"],
      "env": {
        "SCOUTER_COLLECTOR_HOST": "stg-collector", "SCOUTER_COLLECTOR_PORT": "6100",
        "SCOUTER_USER": "stg-user", "SCOUTER_PASSWORD": "***", "SCOUTER_TZ": "Asia/Seoul"
      }
    }
  }
}

The AI then orchestrates across the collectors.

Tools (14)

Fuzzy targeting (objNameLike)

Users say app-name fragments ("shop-order-api"), but real objNames embed the k8s pod name (/shop-order-api-deployment-5f4b8c7d9-abcde/shop-order-api1), so objHash changes on every deploy and an app spans multiple instances. objNameLike solves this: a case-insensitive fragment is resolved to all matching instances (alive first, capped at 20) and queried across them — no objHash needed, ever. For XLog search/summary the resolution also unions the collector's daily object DB, so pods replaced by a deploy during the queried window are still found. If nothing matches, the error is NOT_FOUND with a candidates hint listing actual objNames so the caller can self-correct in one step.

Sloppy service queries

Scouter service names look like /api/order/.../search-order-info-grade<POST>, but users type "GET orderDetail" or "order info grade". The service filter normalizes such input: an HTTP method is extracted from any position (GET x, x POST, pasted <POST>), whitespace-separated words fall back to the longest token server-side, and explicit * patterns pass through untouched. Server-side matching is still case-sensitive — so when a pattern matches nothing, the same window is re-scanned (bounded) without the service filter and real service names matching the query tokens case-insensitively are returned as serviceCandidates, ordered by traffic. One retry with an exact name resolves it.

service/login/ip/desc use substring match by default (server-side StrMatch), so a short token like search-order-info-grade matches /api/order/ext/order-info/search-order-info-grade<POST>. objNameLike/login/ip/desc count as server-side filters, so they relax the 5-minute unfiltered-window cap. list_counters also accepts objNameLike and derives the objType, so users never need to know Scouter's type taxonomy.

All tools are advertised with readOnlyHint. A diagnose_root_cause MCP prompt exposes the recommended tool order for latency/error investigations.

Resource / token safety policy

Production Scouter can produce hundreds of thousands of XLogs in five minutes, so search_xlog enforces guardrails (see scouter.mcp.policy.Limits):

• During streaming, it stops once the limit or the scan cap (5,000 examined packs) is reached and closes the socket, which also stops the Collector's scan/transfer — bounding server load, network, and MCP heap together.
• Without a service or objHash filter, only windows up to 5 minutes are allowed; the absolute window cap is 24 hours.
• limit defaults to 20 and is capped at 200. Results include truncated/scanCapReached and a hint so the caller can narrow filters instead of refetching.
• get_service_summary retains no rows (only per-service counters), so it uses a higher scan cap (200,000) to cover wider windows cheaply; it reports scanCapReached/examined too.
• get_counter caps the per-objType fan-out at 20 instances, and downsamples long series with a min/max scheme that preserves spikes/dips (summary min/max/avg are computed from the full series).
• Windows crossing midnight are split per calendar day (the collector partitions XLogs/counters/alerts by day), so no data is lost on either side of the boundary.
• Response text budgets: SQL text is cut at 1,500 chars, error messages at 500, thread stack traces at 4,000, env values at 500 — each with a truncation marker carrying the original length. get_xlog_detail profile steps are capped at 150, signalled via totalSteps/stepsTruncated.
• A single request may fan out to at most 40 collector round-trips (instances x day segments). When client-side filters (minElapsedMs/onlyError) discard over 99% of scanned rows, a low-selectivity hint steers the model toward server-side filters or get_summary.
• get_summary/get_counter_stat read the collector's daily pre-aggregated data (no scanning), capped at 31 days; summary returns the top 50 rows per category. list_threads caps at 5 alive instances and 50 thread rows each (the state histogram always covers all threads).
• Per-request telemetry (passes/examined/kept/tookMs) is logged to stderr as structured key=value lines for post-hoc load analysis.

Internationalization

Only dynamic, user-facing output (tool error messages, result hints, notes) is localized, in English and Korean, via messages.properties / messages_ko.properties. Static schema/tool descriptions and structured stderr logs (key=value) remain English for a stable contract and log parsing.

Security notes