@thotischner/observability-mcp 1.2.2 → 1.3.1

package/dist/connectors/prometheus.d.ts

@@ -23,6 +23,8 @@ export declare class PrometheusConnector implements ObservabilityConnector {
     listAvailableMetrics(_service: string): Promise<MetricInfo[]>;
     queryMetrics(params: MetricQuery): Promise<MetricResult>;
     private buildQuery;
+    private groupKey;
+    private getDistinctLabelValues;
     private pickMetricCandidate;
     private seriesExistsForService;
     private resolveServiceLabel;

package/dist/connectors/prometheus.js

@@ -3,11 +3,13 @@ const PROMETHEUS_METRIC_CANDIDATES = {
     cpu: [
         {
             seriesName: "process_cpu_seconds_total",
+            // rate() preserves all labels — already broken down per-instance.
             query: 'rate(process_cpu_seconds_total{ {{selector}} }[1m]) * 100',
         },
         {
             seriesName: "node_cpu_seconds_total",
             query: '100 - avg(rate(node_cpu_seconds_total{ {{selector}}, mode="idle" }[1m])) * 100',
+            groupedQuery: '100 - avg by({{groupBy}}) (rate(node_cpu_seconds_total{ {{selector}}, mode="idle" }[1m])) * 100',
         },
     ],
     memory: [
@@ -24,33 +26,40 @@ const PROMETHEUS_METRIC_CANDIDATES = {
         {
             seriesName: "http_requests_total",
             query: 'sum(rate(http_requests_total{ {{selector}} }[1m]))',
+            groupedQuery: 'sum by({{groupBy}}) (rate(http_requests_total{ {{selector}} }[1m]))',
         },
     ],
     error_rate: [
         {
             seriesName: "http_requests_total",
             query: 'sum(rate(http_requests_total{ {{selector}}, status=~"5.." }[1m]))',
+            groupedQuery: 'sum by({{groupBy}}) (rate(http_requests_total{ {{selector}}, status=~"5.." }[1m]))',
         },
     ],
     latency_p99: [
         {
             seriesName: "http_request_duration_seconds_bucket",
             query: 'histogram_quantile(0.99, sum(rate(http_request_duration_seconds_bucket{ {{selector}} }[1m])) by (le))',
+            groupedQuery: 'histogram_quantile(0.99, sum by(le, {{groupBy}}) (rate(http_request_duration_seconds_bucket{ {{selector}} }[1m])))',
         },
     ],
     latency_p50: [
         {
             seriesName: "http_request_duration_seconds_bucket",
             query: 'histogram_quantile(0.50, sum(rate(http_request_duration_seconds_bucket{ {{selector}} }[1m])) by (le))',
+            groupedQuery: 'histogram_quantile(0.50, sum by(le, {{groupBy}}) (rate(http_request_duration_seconds_bucket{ {{selector}} }[1m])))',
         },
     ],
     latency_avg: [
         {
             seriesName: "http_request_duration_seconds_sum",
             query: 'sum(rate(http_request_duration_seconds_sum{ {{selector}} }[1m])) / sum(rate(http_request_duration_seconds_count{ {{selector}} }[1m]))',
+            groupedQuery: 'sum by({{groupBy}}) (rate(http_request_duration_seconds_sum{ {{selector}} }[1m])) / sum by({{groupBy}}) (rate(http_request_duration_seconds_count{ {{selector}} }[1m]))',
         },
     ],
 };
+// Common breakdown labels probed for the auto-hint when no groupBy is set.
+const HINT_BREAKDOWN_LABELS = ["instance", "pod"];
 const DEFAULT_METRIC_META = {
     cpu: { unit: "percent", description: "CPU usage % (auto: prom-client process_cpu_seconds_total or node_exporter node_cpu_seconds_total)" },
     memory: { unit: "bytes", description: "Resident memory bytes (auto: prom-client process_resident_memory_bytes or node_memory used)" },
@@ -187,41 +196,77 @@ export class PrometheusConnector {
         return metrics;
     }
     async queryMetrics(params) {
-        const { promql, label } = await this.buildQuery(params.service, params.metric);
+        const { promql, label, candidate } = await this.buildQuery(params.service, params.metric, params.groupBy);
         const { start, end, step } = this.parseTimeRange(params.duration, params.step);
         const data = await this.apiGet(`/api/v1/query_range?query=${encodeURIComponent(promql)}&start=${start}&end=${end}&step=${step}`);
-        const values = [];
-        const rawValues = [];
-        const resultData = data?.data?.result?.[0]?.values || [];
-        for (const [ts, val] of resultData) {
-            const numVal = parseFloat(val);
-            if (!isNaN(numVal)) {
-                values.push({ timestamp: new Date(ts * 1000).toISOString(), value: numVal });
-                rawValues.push(numVal);
+        const seriesList = data?.data?.result || [];
+        // Build groups from each returned series, keyed either by the explicit
+        // groupBy label (when grouped) or by a synthesized name from any
+        // remaining labels (for naturally per-instance queries like cpu/memory
+        // prom-client). Empty when nothing came back.
+        const groups = [];
+        for (const series of seriesList) {
+            const seriesValues = [];
+            const rawValues = [];
+            for (const [ts, val] of series.values || []) {
+                const numVal = parseFloat(val);
+                if (!isNaN(numVal)) {
+                    seriesValues.push({ timestamp: new Date(ts * 1000).toISOString(), value: numVal });
+                    rawValues.push(numVal);
+                }
             }
+            groups.push({
+                key: this.groupKey(series.metric || {}, params.groupBy),
+                values: seriesValues,
+                summary: this.computeSummary(rawValues),
+            });
         }
-        return {
+        // Top-level values/summary always reflect the first series (back-compat:
+        // single-aggregated queries always return one row, so this is unchanged).
+        const top = groups[0] || { values: [], summary: this.computeSummary([]) };
+        const result = {
             source: this.name,
             service: params.service,
             metric: params.metric,
             unit: this.getUnit(params.metric),
-            values,
-            summary: this.computeSummary(rawValues),
+            values: top.values,
+            summary: top.summary,
             resolvedSeries: promql,
             resolvedLabel: label,
         };
+        if (params.groupBy && groups.length > 1) {
+            result.groupBy = params.groupBy;
+            result.groups = groups;
+        }
+        else if (!params.groupBy && candidate) {
+            // Probe common breakdown labels and hint when more than one distinct
+            // value exists for this service. Helps the model ask the right
+            // follow-up instead of silently looking at an aggregated number.
+            const escaped = params.service.replace(/\\/g, "\\\\").replace(/"/g, '\\"');
+            for (const breakdownLabel of HINT_BREAKDOWN_LABELS) {
+                const distinct = await this.getDistinctLabelValues(candidate.seriesName, label, escaped, breakdownLabel);
+                if (distinct.length > 1) {
+                    result.hint = `${distinct.length} distinct ${breakdownLabel}s exist for this service. Pass groupBy="${breakdownLabel}" to break the result down.`;
+                    break;
+                }
+            }
+        }
+        return result;
     }
     // --- Private helpers ---
-    async buildQuery(service, metric) {
+    async buildQuery(service, metric, groupBy) {
         // Resolve the service-filter label first. Candidate probing uses this
         // label to scope existence checks per-service rather than per-source.
         const escaped = service.replace(/\\/g, "\\\\").replace(/"/g, '\\"');
         let label = "job";
         let template;
+        let candidate = null;
         if (!this.userOverrides.has(metric) && PROMETHEUS_METRIC_CANDIDATES[metric]) {
             label = await this.resolveServiceLabel(service);
-            const candidate = await this.pickMetricCandidate(metric, label, escaped);
-            template = candidate?.query || PROMETHEUS_METRIC_CANDIDATES[metric][0].query;
+            candidate = await this.pickMetricCandidate(metric, label, escaped);
+            const fallback = PROMETHEUS_METRIC_CANDIDATES[metric][0];
+            const chosen = candidate || fallback;
+            template = (groupBy && chosen.groupedQuery) ? chosen.groupedQuery : chosen.query;
         }
         else {
             const def = this.metrics.find((m) => m.name === metric);
@@ -236,8 +281,41 @@ export class PrometheusConnector {
             const selector = `${label}="${escaped}"`;
             promql = promql.replace(/\{\{selector\}\}/g, selector);
         }
+        if (groupBy && template.includes("{{groupBy}}")) {
+            // groupBy is a label name (caller-supplied). Constrain to the same
+            // safe character set we use for service names so it can't break out
+            // of the by(...) clause.
+            const safe = groupBy.replace(/[^a-zA-Z0-9_]/g, "");
+            promql = promql.replace(/\{\{groupBy\}\}/g, safe);
+        }
         promql = promql.replace(/\{\{service\}\}/g, escaped);
-        return { promql, label };
+        return { promql, label, candidate };
+    }
+    groupKey(metric, groupBy) {
+        if (groupBy && metric[groupBy] !== undefined)
+            return metric[groupBy];
+        // No explicit groupBy: synthesize a key from instance/pod/node if any,
+        // else from all labels. Useful for naturally-per-series queries like
+        // process_resident_memory_bytes (no aggregator dropping labels).
+        for (const probe of HINT_BREAKDOWN_LABELS) {
+            if (metric[probe])
+                return metric[probe];
+        }
+        const entries = Object.entries(metric);
+        if (entries.length === 0)
+            return "default";
+        return entries.map(([k, v]) => `${k}=${v}`).join(",");
+    }
+    async getDistinctLabelValues(seriesName, label, escapedService, breakdownLabel) {
+        try {
+            const matchExpr = `${seriesName}{${label}="${escapedService}"}`;
+            const url = `/api/v1/label/${encodeURIComponent(breakdownLabel)}/values?match[]=${encodeURIComponent(matchExpr)}`;
+            const data = await this.apiGet(url);
+            return data?.data || [];
+        }
+        catch {
+            return [];
+        }
     }
     async pickMetricCandidate(metric, label, escapedService) {
         const candidates = PROMETHEUS_METRIC_CANDIDATES[metric];

package/dist/index.js

@@ -57,37 +57,45 @@ async function main() {
     const registry = new ConnectorRegistry();
     await registry.initialize(config);
     applyConfigToRuntime(config, registry);
-    const mcpServer = new McpServer({
-        name: "observability-mcp",
-        version: "1.0.0",
-    });
-    // --- Register tools with Zod schemas ---
-    mcpServer.tool("list_sources", "List all configured observability backends and their connection status. Use this to discover what data sources are available.", {}, async () => listSourcesHandler(registry));
-    mcpServer.tool("list_services", "List all monitored services discovered across all connected backends. Returns service names, their data sources, and signal types (metrics/logs).", { filter: z.string().optional().describe("Optional filter to match service names") }, async (args) => listServicesHandler(registry, args));
-    const metricsList = getAvailableMetricNames(registry);
-    const metricNames = registry.getBySignal("metrics").flatMap(c => c.getMetrics().map(m => m.name));
-    const uniqueNames = [...new Set(metricNames)];
-    mcpServer.tool("query_metrics", `Query a specific metric for a service over a given timeframe. Returns time-series data with pre-computed summary statistics (current, average, min, max, trend). Available metrics: ${metricsList}`, {
-        service: z.string().describe("Service name (e.g. 'api-gateway', 'payment-service')"),
-        metric: z.string().describe(`Metric name. Available: ${uniqueNames.join(", ")}`),
-        duration: z.string().optional().describe("Time range (e.g. '5m', '1h', '24h'). Default: '5m'"),
-        source: z.string().optional().describe("Specific source name. If omitted, queries all metrics backends."),
-    }, async (args) => queryMetricsHandler(registry, args));
-    mcpServer.tool("query_logs", "Query logs for a service over a given timeframe. Returns log entries with a summary including error/warning counts and top error patterns.", {
-        service: z.string().describe("Service name (e.g. 'payment-service')"),
-        query: z.string().optional().describe("Optional search query to filter log messages (regex supported)"),
-        duration: z.string().optional().describe("Time range (e.g. '5m', '1h', '24h'). Default: '5m'"),
-        level: z.string().optional().describe("Filter by log level: 'error', 'warn', 'info', 'debug'"),
-        limit: z.number().optional().describe("Maximum log entries to return. Default: 100"),
-    }, async (args) => queryLogsHandler(registry, args));
-    mcpServer.tool("get_service_health", "Get an aggregated health overview for a service combining metrics AND logs. Returns health score (0-100), status (healthy/degraded/critical), key metrics, log error summary, anomalies, and cross-signal correlations.", {
-        service: z.string().describe("Service name to check health for"),
-    }, async (args) => getServiceHealthHandler(registry, args));
-    mcpServer.tool("detect_anomalies", "Scan for anomalies across all monitored services (or a specific one). Uses z-score analysis on metrics, checks log error spikes, and correlates signals. Returns anomalies with severity ratings.", {
-        service: z.string().optional().describe("Specific service to scan. If omitted, scans all."),
-        duration: z.string().optional().describe("Time range to analyze (e.g. '5m', '15m', '1h'). Default: '10m'"),
-        sensitivity: z.enum(["low", "medium", "high"]).optional().describe("Detection sensitivity: low (>3σ), medium (>2σ), high (>1.5σ). Default: 'medium'"),
-    }, async (args) => detectAnomaliesHandler(registry, args));
+    // The MCP SDK Protocol class permits exactly one transport per instance,
+    // so we cannot share a single McpServer across HTTP sessions. Each new
+    // session needs its own server. The factory captures the live registry
+    // by reference so tool handlers always see the current configuration.
+    function createMcpServer() {
+        const mcpServer = new McpServer({
+            name: "observability-mcp",
+            version: "1.3.0",
+        });
+        // --- Register tools with Zod schemas ---
+        mcpServer.tool("list_sources", "List all configured observability backends and their connection status. Use this to discover what data sources are available.", {}, async () => listSourcesHandler(registry));
+        mcpServer.tool("list_services", "List all monitored services discovered across all connected backends. Returns service names, their data sources, and signal types (metrics/logs).", { filter: z.string().optional().describe("Optional filter to match service names") }, async (args) => listServicesHandler(registry, args));
+        const metricsList = getAvailableMetricNames(registry);
+        const metricNames = registry.getBySignal("metrics").flatMap(c => c.getMetrics().map(m => m.name));
+        const uniqueNames = [...new Set(metricNames)];
+        mcpServer.tool("query_metrics", `Query a specific metric for a service over a given timeframe. Returns time-series data with pre-computed summary statistics (current, average, min, max, trend). Available metrics: ${metricsList}`, {
+            service: z.string().describe("Service name (e.g. 'api-gateway', 'payment-service')"),
+            metric: z.string().describe(`Metric name. Available: ${uniqueNames.join(", ")}`),
+            duration: z.string().optional().describe("Time range (e.g. '5m', '1h', '24h'). Default: '5m'"),
+            source: z.string().optional().describe("Specific source name. If omitted, queries all metrics backends."),
+            groupBy: z.string().optional().describe("Label to break the result down by, e.g. 'instance', 'pod', 'node'. Returns one series per distinct value in 'groups'."),
+        }, async (args) => queryMetricsHandler(registry, args));
+        mcpServer.tool("query_logs", "Query logs for a service over a given timeframe. Returns log entries with a summary including error/warning counts and top error patterns.", {
+            service: z.string().describe("Service name (e.g. 'payment-service')"),
+            query: z.string().optional().describe("Optional search query to filter log messages (regex supported)"),
+            duration: z.string().optional().describe("Time range (e.g. '5m', '1h', '24h'). Default: '5m'"),
+            level: z.string().optional().describe("Filter by log level: 'error', 'warn', 'info', 'debug'"),
+            limit: z.number().optional().describe("Maximum log entries to return. Default: 100"),
+        }, async (args) => queryLogsHandler(registry, args));
+        mcpServer.tool("get_service_health", "Get an aggregated health overview for a service combining metrics AND logs. Returns health score (0-100), status (healthy/degraded/critical), key metrics, log error summary, anomalies, and cross-signal correlations.", {
+            service: z.string().describe("Service name to check health for"),
+        }, async (args) => getServiceHealthHandler(registry, args));
+        mcpServer.tool("detect_anomalies", "Scan for anomalies across all monitored services (or a specific one). Uses z-score analysis on metrics, checks log error spikes, and correlates signals. Returns anomalies with severity ratings.", {
+            service: z.string().optional().describe("Specific service to scan. If omitted, scans all."),
+            duration: z.string().optional().describe("Time range to analyze (e.g. '5m', '15m', '1h'). Default: '10m'"),
+            sensitivity: z.enum(["low", "medium", "high"]).optional().describe("Detection sensitivity: low (>3σ), medium (>2σ), high (>1.5σ). Default: 'medium'"),
+        }, async (args) => detectAnomaliesHandler(registry, args));
+        return mcpServer;
+    }
     // --- HTTP server ---
     const app = express();
     app.use(express.json({ limit: "1mb" }));
@@ -378,7 +386,8 @@ async function main() {
                     }
                 }
             };
-            await mcpServer.connect(transport);
+            const sessionMcpServer = createMcpServer();
+            await sessionMcpServer.connect(transport);
         }
         await transport.handleRequest(req, res, req.body);
         // Store session after handling (sessionId is set during handleRequest)

package/dist/tools/query-metrics.d.ts

@@ -21,6 +21,10 @@ export declare const queryMetricsDefinition: {
                 type: string;
                 description: string;
             };
+            groupBy: {
+                type: string;
+                description: string;
+            };
         };
         required: string[];
     };
@@ -30,6 +34,7 @@ export declare function queryMetricsHandler(registry: ConnectorRegistry, args: {
     metric: string;
     duration?: string;
     source?: string;
+    groupBy?: string;
 }): Promise<{
     content: {
         type: "text";

package/dist/tools/query-metrics.js

@@ -21,6 +21,10 @@ export const queryMetricsDefinition = {
                 type: "string",
                 description: "Specific source name to query. If omitted, queries all metrics backends.",
             },
+            groupBy: {
+                type: "string",
+                description: "Optional label to break the result down by, e.g. 'instance', 'pod', 'node'. When set, the response includes a 'groups' array with one time-series per distinct value. When the service has only one matching series, the result is unchanged.",
+            },
         },
         required: ["service", "metric"],
     },
@@ -36,6 +40,9 @@ export async function queryMetricsHandler(registry, args) {
     const metricErr = validateMetricName(args.metric, registry);
     if (metricErr)
         return errorResponse(metricErr);
+    if (args.groupBy && !/^[a-zA-Z_][a-zA-Z0-9_]*$/.test(args.groupBy)) {
+        return errorResponse(`Invalid groupBy "${args.groupBy}". Must be a valid Prometheus label name (alphanumeric + underscore, starting with letter/underscore).`);
+    }
     const connectors = args.source
         ? [registry.getByName(args.source)].filter(Boolean)
         : registry.getBySignal("metrics");
@@ -55,6 +62,7 @@ export async function queryMetricsHandler(registry, args) {
                 service: args.service,
                 metric: args.metric,
                 duration,
+                groupBy: args.groupBy,
             });
             results.push(result);
         }

package/dist/types.d.ts

@@ -95,6 +95,7 @@ export interface MetricQuery {
     metric: string;
     duration: string;
     step?: string;
+    groupBy?: string;
 }
 export interface LogQuery {
     service: string;
@@ -114,6 +115,11 @@ export interface MetricSummary {
     max: number;
     trend: Trend;
 }
+export interface MetricGroup {
+    key: string;
+    values: DataPoint[];
+    summary: MetricSummary;
+}
 export interface MetricResult {
     source: string;
     service: string;
@@ -123,6 +129,9 @@ export interface MetricResult {
     summary: MetricSummary;
     resolvedSeries?: string;
     resolvedLabel?: string;
+    groupBy?: string;
+    groups?: MetricGroup[];
+    hint?: string;
 }
 export interface LogEntry {
     timestamp: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@thotischner/observability-mcp",
-  "version": "1.2.2",
+  "version": "1.3.1",
   "description": "Unified observability gateway for AI agents — one MCP server for Prometheus, Loki, and any backend",
   "type": "module",
   "license": "MIT",