multi-agent-credit-desk

multi-agent-credit-desk is described as a multi-agent platform that analyzes corporate credit applications for Brazilian legal entities (pessoas jurídicas, PJ) and produces an auditable credit decision with reproducible evidence. It is built on Python 3.13, uv, and the team Claude Code engineering harness. The repository states that all customer and Open Finance data used or planned for the project is synthetic, and no real Open Finance, BCB, or credit bureau connection exists or is planned without an explicit, separately reviewed integration.

The workspace contains packages/contracts, which provides versioned Pydantic v2 schemas for artifact envelopes, structured events, and model-routing decisions, and packages/credit-core, which implements a deterministic, policy-driven credit evaluation core (score, blocking rules, approval-authority policy) behind a synthetic demo policy. services/policy-mcp is a read-only MCP server that exposes a versioned catalog of the credit policy credit_core enforces. services/bureau-mcp is a read-only MCP server that exposes a synthetic credit-bureau report (external score, negative records) for a fixed set of demo companies, keyed by CNPJ.

services/decisao-agent calls credit_core.evaluation.evaluate_credit_application directly and cross-checks the result against policy-mcp's catalog over the real MCP protocol. It is also the workspace's first real A2A agent, advertising one skill, evaluate_credit_application, over the official a2a-sdk. services/cadastral-agent screens a company's KYC standing by fetching its bureau-mcp report and applying a small, deterministic policy over the external score and adverse records, producing APPROVED, COMMITTEE_REFERRAL, or BLOCKED. The repository notes that no orchestrator is implemented yet, and risco-agent is not implemented in this repository yet.