Loading ToolBench
AgentUtils operations over MCP. Remote MCP server that exposes operations from a shared registry as tools, with actor-scoped access control and HTTP envelope responses.
Critical failures across all dimensions. Tool definitions are inferred from file references only, no explicit tool registration visible in source. Four tools listed (kv-store, audit-log, dlq, checkpoint) but source code provided shows only Next.js/Dockerfile configuration, analytics components, and doc pages, NO actual MCP tool implementations found. Descriptions are present but lack actionable detail (e.g., 'Human-in-the-Loop. Pause agents until humans approve.' is 8 words, far below 20-char minimum). Input schemas visible for kv-store and audit-log but incomplete: kv-store lacks 'cas_token' parameter mentioned in description; audit-log and dlq lack output schemas entirely. checkpoint has no input schema visible. No error handling guidance, no parameter constraints (enums, ranges), no recovery paths. Tool composition unclear, kv-store and checkpoint appear to serve overlapping state-management concerns without clear separation.
Audit Logging for AI Agents. Immutable records. Server timestamps. Compliance-ready. Append immutable audit entries that record what your agent did, when, and why. Every entry is server-timestamped and stored append-only.
Human-in-the-Loop. Pause agents until humans approve.
Dead Letter Queue for AI Agents. Catch failed tasks. Inspect payloads. Retry with one click. When an agent task fails, catch the error payload and original input in a dead letter queue. Inspect failures, debug payloads, and retry.
Key-Value Store for AI Agent State. Per-agent state. Compare-and-set. No database to run. Give every agent its own namespaced key-value store. Read and write JSON values with optional TTL expiry and compare-and-set (CAS) tokens for safe concurrent updates. State is isolated per agent and per tenant.
Tool implementations not found in source. Only file references (lib/seo-tools.ts) provided; actual tool registration code absent. Cannot verify schemas, error handling, or composition.
checkpoint description is 8 words ('Human-in-the-Loop. Pause agents until humans approve.'), far below 20-char minimum. Fails to explain WHAT it does, WHEN to use it, or what it returns. No input schema visible.
kv-store description mentions 'compare-and-set (CAS) tokens' but input schema has no cas_token parameter. Mismatch between documented behavior and schema. Output schema not documented.
audit-log and dlq lack output schemas entirely. LLMs cannot plan downstream calls or extract required fields. No pagination, no result limits documented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
No error handling guidance across any tool. No recovery paths, no categorization of retryable vs fatal errors, no actionable error messages. Agents cannot self-correct on failure.