Loading ToolBench
Model Context Protocol server for ARGUS fraud detection platform. Provides database tools for transaction analysis, risk scoring, and fraud case management via Lambda function URL with SigV4 authentication.
ARGUS has 9 well-defined tools with complete input schemas and descriptions. All tools follow verb_noun naming (get_*, update_*, log_*). Descriptions are present and actionable (avg ~100 chars). However, parameter descriptions are sparse or missing entirely, most params lack guidance on format, constraints, or valid ranges. Output schemas are not documented. Error handling is minimal (generic 'not found' responses). No tool annotations (readOnlyHint/destructiveHint). The two WRITE tools (update_case_verdict, log_agent_action) lack confirmation or dry-run patterns. Overall: solid foundation, but parameter-level documentation and output schema clarity are weak.
Get the full account profile of whoever made a given transaction.
Fetch recent transaction history for an account to spot behavioural patterns.
List open fraud cases ordered by risk score.
Retrieve all risk signals and rule violations fired against a transaction.
Find other transactions with similar IP, device, or amount patterns.
Retrieve full details of a transaction including account and merchant information.
Parameter descriptions missing or incomplete. Most params (e.g., txn_id, account_id, verdict, reasoning) lack format guidance, constraints, or examples. LLMs cannot infer whether txn_id is 'TXN-20240401-8821' or a UUID.
Output schemas not documented. Tools return lists or dicts but LLMs have no schema to understand field names, types, or structure. E.g., get_transaction returns a dict with 20+ fields, no schema tells the LLM what to expect.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
Append an entry to the agent audit log.
Write AI verdict back to the fraud case.
Count how many transactions an account made in a rolling time window. Detects velocity fraud and burst patterns.
No tool annotations. WRITE tools (update_case_verdict, log_agent_action) lack destructiveHint=true. READ tools lack readOnlyHint=true. Agents cannot distinguish safe from unsafe operations without explicit hints.
Destructive operations lack confirmation. update_case_verdict modifies fraud case status and risk_score without dry-run or confirmation. Agents can escalate cases or change verdicts without review.
Error responses are generic. 'Transaction not found' or 'No account found' provide no recovery guidance. Should suggest: 'Try search_transactions() with a partial merchant name' or 'Available accounts: [list]'.