convez.ai
MCP ServerFinanceiroAvançado
zavora-ai

Credit Bureau MCP Server

zavora-ai/mcp-credit-bureau

Bureau de crédito regulado com FCRA, scores e análise de affordability

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP em Rust que implementa uma plataforma de bureau de crédito com conformidade FCRA, oferecendo pull de relatórios de crédito com propósito permitido, cálculo de scores transparentes, análise de affordability e registro de acesso. Integra-se a agentes de decisão bancária para análise de crédito, revisão de limites e underwriting.

Instalação

$npx convez-ai zavora-ai/mcp-credit-bureau

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
0
Acompanhando
0
Saúde
50%
Último commit
há 3 meses
Versão
v1.1.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Underwriting de empréstimos com análise de tradelines e histórico de pagamentos
  • Revisão de limites de crédito com cálculo de DTI e affordability
  • Disputes FCRA e monitoramento de alertas de crédito
  • Logging de acesso regulado para conformidade com FCRA

Veredito

Solução altamente especializada para banking agents que precisam integrar decisões de crédito com conformidade regulatória. Requer profunda compreensão de FCRA e arquitetura de decisão creditícia; não é para uso geral.

README

A tradução para o português deste README ainda está na fila.

Credit Bureau MCP Server

Crates.io License ADK-Rust Enterprise Registry Ready

A credit-reporting (credit bureau) platform for ADK-Rust Enterprise banking agents. 14 MCP tools covering consumers, credit reports (tradelines, inquiries, public records), scores with reason factors, permissible-purpose-gated report pulls that record a hard inquiry, FCRA disputes, monitoring alerts, and affordability/decision analytics — with an FCRA-style access log on every consumer-scoped access.

⚠️ Regulated, PII-sensitive domain. Accessing a consumer report requires a permissible purpose; every access is logged. This is a reference engine with fictitious data — see FCRA posture.

A platform, not a point solution

This is modeled as a general credit-bureau backbone (the layer behind a CRA / bureau access API), so banking decisioning agents are clients of one shared system:

AgentDomainUses
Loan Underwriting Assistantbankingpull_credit_report (credit_application = hard), get_score, affordability
Credit Limit Review Agentbankingpull_credit_report (account_review = soft), get_score, affordability, alerts

Architecture

Capabilities

  • Consumers — fictitious identities (SSN as a reference handle, minimal address PII) with a security freeze flag.
  • Credit file — tradelines (revolving/installment/mortgage/auto/student) with balances, limits, payments, lates; public records (bankruptcy/lien/judgment); inquiries.
  • Scores — a transparent 300–850 rule-based model (utilization, delinquencies, derogatory accounts, public records, hard inquiries, thin file, age) with adverse-action reason factors. (Reference model, not FICO/VantageScore.)
  • Report pulls — pull_credit_report assembles the file under a stated permissible purpose; credit_application is a hard inquiry (affects score), others are soft.
  • Affordability — utilization, monthly obligations, and DTI vs a stated income (healthy/elevated/high) for underwriting and limit review.
  • Disputes — FCRA disputes flag the tradeline and set a ~30-day reinvestigation due date; resolution clears the flag.
  • Monitoring — alerts (new inquiry, delinquency, etc.).

FCRA posture

  • Permissible purpose required — pull_credit_report takes a PermissiblePurpose and records it. New-credit applications are hard inquiries; account review / collection / consumer disclosure / employment are soft.
  • Accounting of disclosures — every consumer-scoped access (identity read, search, pull, score, affordability, alerts, disputes) is written to a per-consumer access log (access_log); consumer-scoped tools require an actor.
  • Security freeze — a frozen consumer's new-credit (hard) pulls are blocked until the freeze is lifted; soft account-review pulls remain allowed.
  • Consumer rights — file_dispute / resolve_dispute model FCRA reinvestigation; set_freeze models the freeze right.
  • Gated, high-impact writes — pull_credit_report (external_write, records a hard inquiry affecting the score) and resolve_dispute (changes what's reported) require approval.
  • Reference engine — fictitious data; SSNs are handles. A production deployment must enforce FCRA, GLBA, ECOA/Reg B, and adverse-action notice requirements with real controls, identity proofing, and audited access — software gates here are defense-in-depth, not compliance.

Tools (14)

Consumers (4)

create_consumer · get_consumer · find_consumers · set_freeze

Credit File (2)

add_tradeline · add_public_record

Reports & Scoring (3)

pull_credit_report (gated, external, permissible-purpose) · get_score · affordability

Disputes, Monitoring & Disclosure (5)

file_dispute · resolve_dispute (gated) · list_disputes · alerts · access_log

Example

// Underwriting: a permissible-purpose hard pull, then decision analytics
{"name": "pull_credit_report", "arguments": {"actor": "uw.agent", "consumer_id": "CON-1001",
  "purpose": "credit_application", "requester": "BankA Underwriting"}}
{"name": "affordability", "arguments": {"actor": "uw.agent", "consumer_id": "CON-1001", "monthly_income": 6000}}

// FCRA: dispute + accounting of disclosures
{"name": "file_dispute", "arguments": {"consumer_id": "CON-1004", "tradeline_id": "TRL-1005", "reason": "not mine"}}
{"name": "access_log", "arguments": {"actor": "examiner", "consumer_id": "CON-1001"}}

Install & run

cargo install mcp-credit-bureau
mcp-credit-bureau            # serves MCP over stdio

Or build from source:

git clone https://github.com/zavora-ai/mcp-credit-bureau
cd mcp-credit-bureau && cargo build --release
./target/release/mcp-credit-bureau

Registry manifest

server_id = "mcp_credit_bureau"
display_name = "Credit Bureau"
version = "1.0.0"
domain = "banking"
risk_level = "high"
writes_allowed = "gated"

The full mcp-server.toml declares all 14 tools with risk classes and approval gates for registry onboarding.

License

Apache-2.0

Ver o original no GitHub

Bom para

Relacionados