Loading ToolBench
Exposes Sentinel runtime queries and actions as MCP tools for querying and controlling Sentinel from any MCP client (Claude Code, Cursor, etc.)
EVAV Sentinel MCP exposes 11 tools with explicit schemas and descriptions, but quality is inconsistent. Naming follows verb_noun convention (sentinel_*), which is good. However, parameter descriptions are largely absent, most params lack explanations of what they control or their constraints. Enum values are declared but not always documented (e.g., 'severity' enum values listed but not explained). Output schemas are not documented anywhere, forcing LLMs to infer response structure. Error handling is minimal, no recovery guidance or actionable error messages visible. Two tools (sentinel_fire_intervention, sentinel_judge_score) have enum constraints with null values, indicating incomplete schema definitions. Overall, the server provides basic structure but lacks the depth needed for confident LLM reasoning.
Current behavior baselines for the authenticated customer
Recent federated alerts across the EVAV network
Manually fire an intervention. Returns intervention_id.
Record outcome of an intervention (success/failure/unknown). Feeds the bandit.
Run LLM-as-judge rubrics on an event
Pending novelty signals waiting for review
Parameter descriptions missing. Most parameters (limit, severity, agent_id, reason, rule_violated, etc.) lack descriptions explaining what they control or their constraints. LLMs cannot infer parameter meaning from names alone.
Output schemas not documented. No tool documents what fields it returns or their types. LLMs must guess response structure, risking failed downstream tool calls and context loss.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | <=2025-11-25 | v2 |
Use the bandit to pick the best intervention for a rule (without firing)
Get recent events filtered by severity/agent
Get recent interventions fired
Replay events for a given trace_id in chronological order
Get integrations + event/intervention counts for the day
Enum constraints with null values. sentinel_fire_intervention and sentinel_judge_score declare enum constraints but set them to null, indicating incomplete schema definitions. This breaks tool invocation.
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. A failed intervention fire returns nothing actionable.
No pagination or result limits documented. Tools like sentinel_recent_events accept a 'limit' parameter but do not document the maximum allowed value, default behavior, or whether pagination is supported.