Observability and monitoring MCP server that exposes LogSentinel's alert management, log retrieval, and error rate statistics as MCP tools by proxying FastAPI endpoints over HTTP. Includes LangGraph-based monitoring agent for intelligent alert decisions and root-cause analysis.
LogSentinel exposes 5 tools via FastMCP with basic HTTP transport. All tools have descriptions (10-50 chars) and input schemas are present. However, descriptions are minimal and lack LLM-optimization guidance (pattern:tool-description baseline 194 chars). Parameter descriptions exist but are terse. No output schemas documented. Error handling returns raw dicts without recovery guidance. Tools are read-only except trigger_alert_check, but no destructiveHint annotations present. No tool composition guidance or chaining IDs in responses.
List currently active alerts across all services.
Return recent alert history (most recent first).
Return the current (1h window) error rate for a given service.
Return the most recent logs for a given service.
Manually run the monitoring engine and return any services it triggered.
Descriptions are 10-50 characters, well below the 194-char baseline for A+ tools. Lack context on WHEN to use each tool and what it returns. 'List currently active alerts across all services' does not explain when an LLM should call this vs get_alert_history.
No output schemas documented. LLMs cannot infer what fields to expect from responses (e.g., does get_active_alerts return {alerts: [...]} or [{id, severity, service, ...}]?). Breaks tool chaining and forces agents to guess field names.
Error handling returns raw dicts like {"error": "HTTP 500", "detail": "..."} without recovery guidance. Pattern:recovery-guide requires actionable next steps. E.g., 'Service not found. Try get_active_alerts() to see available services.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
trigger_alert_check is a WRITE operation (modifies monitoring state) but has no destructiveHint annotation and no confirmation/dry-run pattern. Agents may invoke it unintentionally.
Parameter descriptions are minimal (e.g., 'Maximum number of log records to return' for limit). Missing constraints: no min/max bounds stated, no format guidance. Baseline is 72 chars; these are 30-50 chars.