The Mifos MCP (Model Context Protocol) project provides an MCP implementation for the Mifos X ecosystem, enabling AI agents to access financial data and perform operations.
Purpose The project serves as an independent, stateless integration layer that bridges any AI assistant or agent framework to the Apache Fineract banking backend. The repository itself is framework-agnostic; the client (LLM brain, UI, memory) resides in another repository, and any MCP-compatible system can connect.
Key Capabilities
- Four implementations: Go (native) provides 102 typed tools, with high performance, cloud-native, and SSE/Stdio; Java (Quarkus) provides 38 typed tools, divided into Backoffice and Recommendations; Python (FastMCP) provides 49 typed tools, using a modular domain-driven design; Rust provides 89 typed tools, with high-performance async I/O and exclusive batch operations.
- Tool coverage: customers and identities, documents and reports, loans and savings, groups and centers, batch and composite operations, accounting and statistics, and other core banking operations. The number and categories of tools differ across implementations; Go has the most complete tools, Rust has exclusive batch operations, and Java has exclusive user recommendation tools.
- All implementations are based on the same Apache Fineract REST API and share field routing logic.
- Strictly stateless: does not store user data, PII, or AI memory.
- Shared smoke tests ensure that all implementations return consistent JSON structures.
Usage Prepare the environment for the chosen language and build and run according to the documentation. Go: requires Go 1.21+; copy go/.env.example to go/.env and update the credentials; run cd go && go build -o mcp-server . && ./mcp-server; you can set PORT=8080 in .env to switch to SSE microservice mode. Rust: requires Rust (Cargo); copy rust/.env.example to rust/.env and update the credentials; run cd rust && cargo build --release && ./target/release/mcp-rust-mifosx. Java: requires JDK 21+, Maven; set the environment variables MIFOSX_BASE_URL, MIFOSX_BASIC_AUTH_TOKEN, MIFOS_TENANT_ID; it can be run via JBang, or enter java/backoffice to build a native executable. Python: requires Python 3.10+, pip; enter the python directory, install the dependencies from requirements.txt, copy .env.example to .env and fill it in, and run python mcp_server.py. You can use MCP Inspector to interactively test and debug the server.
Applicable Scenarios This architecture describes an agent platform: user channels include Mifos WebApp (chat/Copilot UI), Mifos Pay, Mifos Mobile, Mifos Field Officer, and Mifos Web Banking, with Agent CLI as an optional entry point; the Agentic Gateway reuses the existing Mifos AI Gateway endpoints to route requests to the agent runtime; the runtime (Pi / Prime / Hermes) encapsulates the LLM provider, LLM firewall, observability, and session management for each agent; specialized agents include Enrollment Agent, KYC/FT Reviewer, Loan Reviewer, Scoring Reviewer, Document Reviewer, etc. Its Fineract tool calls use the credentials and permissions of the calling officer; agents do not have a separate Fineract role.
Limitations All implementations follow a stateless design and do not store user data or AI memory; the server does not make external calls; every operation is RBAC-validated according to native Fineract permissions, and if the calling officer's role lacks the required authorization, the request is denied. All implementations maintain functional parity, but the number of tools and features differ.
