The verified intelligence layer for AI agents — 62 MCP tools, multi-agent swarm, enterprise safety
Mixed quality across 5 tools. All tools have descriptions and input schemas present, which is good. However, descriptions vary in quality and completeness. Tool naming follows verb_noun convention (veroq_ask, veroq_verify, veroq_analyze_ticker, veroq_verify_market_claim, veroq_generate_trading_signal), which is solid. Parameters are typed with Zod schemas. However, several critical issues: (1) descriptions are inconsistently detailed, some are 200+ chars with context and WHEN TO USE guidance, others are under 100 chars without usage guidance; (2) output schemas are NOT documented anywhere, there is NO type definition for what these tools return beyond vague prose descriptions like 'returns structured data...'; (3) error handling is not visible in the source code, no guidance on what happens when claims are unverifiable, API calls fail, or tickers don't exist; (4) no pagination or result limits documented for tools returning lists/collections; (5) security: no audit logging, no permission gate patterns, no rate limit guidance visible.
Complete analysis of any stock, ETF, or crypto ticker in one call. Price with change %, 20 technical indicators with buy/sell signal, next earnings date, sentiment direction, insider transactions, analyst consensus, and top 5 headlines. Replaces chaining veroq_ticker_price → veroq_technicals → veroq_earnings → veroq_ticker_news.
The most important tool — ask any question in natural language and get verified intelligence. Automatically detects 41 intents (price, technicals, earnings, sentiment, screener, backtest, competitors, insider, filings, analysts, congress, crypto, forex, economy, and more) and routes to the right data sources. Returns structured data from all matched endpoints + LLM-generated natural language summary + composite trade signal (0-100) + confidence level (high/medium/low) + follow-up suggestions.
Generate actionable trading signals by screening stocks and analyzing technicals. Describe what you're looking for in natural language — the NLP screener translates it into filters (sector, RSI, sentiment, volume). Returns matched tickers with price, RSI, sentiment score, and a composite trade signal (0-100) combining RSI (30%), sentiment (25%), VIX (20%), technicals (15%), and earnings proximity (10%).
Fact-check any claim with full evidence chain, confidence breakdown, and source reliability scores. After any agent makes a factual claim about earnings, revenue, market movements, mergers, acquisitions, or any financial data. Returns verdict (supported/contradicted/partially_supported/unverifiable), confidence (0-1) with 4-factor breakdown (source_agreement, source_quality, recency, corroboration_depth), evidence_chain with source names, direct quotes, and URLs, plus a verification receipt.
Output schemas completely undocumented. Tools return objects with fields like 'summary', 'trade_signal', 'confidence', 'evidence_chain', 'verdict', 'confidence_breakdown', all described in prose only, never in a formal schema. LLMs cannot reliably extract structured output without knowing field types and nesting. This violates pattern:tool requirement that output schemas be documented.
No error handling guidance. The source shows no handling for: unverifiable claims, missing tickers, API failures, rate limiting, or malformed input. Error responses are not documented. Per pattern:recovery-guide, error responses must tell the LLM what to do next, e.g. 'Ticker XYZ not found. Try search_tickers() or check symbol validity.' Without this, agents cannot recover from failures.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2025-06-18+ | v2 |
Fact-check any financial or market claim with full evidence chain. Returns verdict (supported/contradicted/partially_supported/unverifiable), confidence score with 4-factor breakdown (source_agreement, source_quality, recency, corroboration_depth), evidence chain with source names, direct quotes, and URLs, plus a verification receipt. Checks corpus first, falls back to live web search.
Pagination not implemented or documented. Tools like veroq_generate_trading_signal return lists (matched tickers) but no pagination, offset, or limit enforcement is visible. Tool docs mention 'max results (1-50, default 20)' for veroq_generate_trading_signal only. Others are silent. Per pattern:paginated-result, list-returning tools must support pagination to avoid context window explosion.
Description completeness varies. veroq_analyze_ticker has rich description with WHEN TO USE and COST guidance (82 chars baseline threshold violated, description is ~380 chars, well above the 10-1024 guideline but within reason). However, veroq_verify has a shorter description lacking WHEN TO USE clarity (70 chars). Per pattern:tool-description baseline, all descriptions should include WHAT, WHEN, and any prerequisites. Inconsistency makes some tools less discoverable.
Parameter descriptions lack constraint details. For example, veroq_ask 'question' param says 'be specific for best results' but does not specify min/max length, character restrictions, or what makes a question 'specific.' Per pattern:constrained-input, descriptions should include format, range, and examples. Current prose is vague and forces LLMs to guess.
No audit logging, permission gates, or scope declarations visible. Tools manipulate financial data and generate trading signals but no pattern:audit-trail, pattern:permission-gate, or pattern:scope-declaration implementation is evident. LLMs should declare what permissions they require (read:market_data, write:signals). No per-tool permission checks visible.
No idempotency guarantees. For tools like veroq_verify that perform fact-checking and searches, it is unclear whether repeated calls with identical input always return identical results. Per pattern:idempotent-operation, agents retry on ambiguous failures, non-idempotent tools risk duplicate side effects. This should be documented for each tool.