Model Context Protocol (MCP) server that reconciles ISO 20022 camt.053 bank-statement entries against expected pain.001 payments -- explainable exact, partial, one-to-many and many-to-one matching.
Strong domain-specific tool set with clear, detailed descriptions and well-structured schemas. All 10 tools have descriptions (avg ~180 chars, within baseline 194). Input schemas are comprehensive with type declarations and parameter descriptions. Tool names follow verb_noun pattern (reconcile, explain_match, normalize_*, list_*, load_*, run_*, match_*). However, output schemas are not explicitly documented in the visible code, only inferred from descriptions. Error handling guidance is minimal. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) visible despite all tools being READ_ONLY. Parameter descriptions are strong but some lack explicit constraints (e.g., 'threshold' in match_names_probabilistic defaults to 0.85 but range [0,1] is stated only in description, not schema).
Score a single expected/observed pair and break down every signal (reference, amount, date, name). A tuning aid -- it explains the score even for pairs below the review threshold.
List the built-in sandbox scenarios (test-mode fixtures). Each demonstrates one reconciliation outcome so you can try the flow with zero real data.
Return the expected/observed inputs for one named sandbox scenario, so you can inspect or edit the fixture before reconciling.
Compare two amounts in different currencies, converting one via a supplied FX rate and matching if the percentage difference is within tolerance. Uses exact decimal arithmetic.
Score two counterparty names with Jaro-Winkler similarity and report whether they match at a given threshold. Tolerant of legal-suffix drift, e.g. 'ACME Corp' vs 'ACME Corporation Inc'.
Convert parsed camt.053 statement entries into canonical observed records ready to reconcile. Accepts a list of entries or a dict wrapping them under 'entries'/'transactions'/'statements'.
Output schemas not explicitly documented. Tool descriptions infer return structure (e.g., 'returns exact matches, short/over payments...') but no formal JSON Schema for response objects visible in code. LLMs cannot reliably plan downstream tool calls without knowing response field names and types.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) not visible in tool definitions. All 10 tools are READ_ONLY per the feature list, but this is not declared in the schema. Agents cannot distinguish safe-to-retry tools from stateful ones without explicit hints.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Convert parsed pain.001 payment instructions into canonical expected records ready to reconcile. Accepts a list of transactions or a dict wrapping them under 'transactions'/'payments'/'records'.
Reconcile expected payments against observed bank-statement entries, returning exact matches, short/over payments, split settlements (one-to-many), batch credits (many-to-one) and unmatched residuals, each with an explainable score and reasons.
Reconcile expected payments against observed entries using integer linear programming to find the globally optimal many-to-many matching. Requires scipy (install via the 'ilp' extra). Falls back gracefully if scipy is absent.
Load a named sandbox scenario and immediately reconcile it -- the one-call way to see a full, explainable result with zero setup. Great for a first run or a smoke test.
Numeric parameter constraints not formalized in schema. 'threshold' in match_names_probabilistic has range [0,1] stated only in description text, not as JSON Schema min/max. 'tolerance_percent' in match_amounts_with_fx_drift similarly lacks schema bounds. LLMs cannot parse description text for constraints reliably.
Error handling guidance absent. No visible error responses or recovery hints. E.g., if reconcile() receives malformed records or normalize_pain001() fails to parse, what should the agent do? No actionable error messages or retry guidance documented.
Pagination not applicable to most tools, but list_sandbox_scenarios() returns an unbounded list. No limit parameter or pagination guidance visible. If scenarios grow, response could bloat context window.