fi-mcp-dev

A minimal, hackathon-ready version of the Fi MCP server. This project provides a lightweight mock server for use in hackathons, demos, and development, simulating the core features of the production Fi MCP server with dummy data and simplified authentication.

Purpose

• fi-mcp-dev is designed for hackathon participants and developers who want to experiment with the Fi MCP API without accessing real user data or production systems.
• It serves dummy financial data and uses a dummy authentication flow, making it safe and easy to use in non-production environments.

Features

• Simulates Fi MCP API: Implements endpoints for net worth, credit report, EPF details, mutual fund transactions, and bank transactions.
• Dummy Data: All responses are served from static JSON files in test_data_dir/, representing various user scenarios.
• Dummy Authentication: Simple login flow using allowed phone numbers (directory names in test_data_dir/). No real OTP or user verification.
• Hackathon-Ready: No real integrations, no sensitive data, and easy to reset or extend.

Directory Structure

• main.go — Entrypoint, sets up the server and endpoints.
• middlewares/auth.go — Implements dummy authentication and session management.
• test_data_dir/ — Contains directories named after allowed phone numbers. Each directory holds JSON files for different API responses (e.g., fetch_net_worth.json).
• static/ — HTML files for the login and login-successful pages.

Dummy Data Scenarios

The dummy data covers a variety of user states. Example scenarios:

• All assets connected: Banks, EPF, Indian stocks, US stocks, credit report, large or small mutual fund portfolios.
• All assets except bank account: No bank account, but other assets present.
• Multiple banks and UANs: Multiple bank accounts and EPF UANs, partial transaction coverage.
• No assets connected: Only a savings account balance is present.
• No credit report: All assets except credit report.

Test Data Scenarios

Example: Dummy Data File

A sample fetch_net_worth.json (truncated for brevity):

{
  "netWorthResponse": {
    "assetValues": [
      {"netWorthAttribute": "ASSET_TYPE_MUTUAL_FUND", "value": {"currencyCode": "INR", "units": "84642"}},
      {"netWorthAttribute": "ASSET_TYPE_EPF", "value": {"currencyCode": "INR", "units": "211111"}}
    ],
    "liabilityValues": [
      {"netWorthAttribute": "LIABILITY_TYPE_VEHICLE_LOAN", "value": {"currencyCode": "INR", "units": "5000"}}
    ],
    "totalNetWorthValue": {"currencyCode": "INR", "units": "658305"}
  }
}

Authentication Flow

• When a tool/API is called, the server checks for a valid session.
• If not authenticated, the user is prompted to log in via a web page (/mockWebPage?sessionId=...).
• Enter any allowed phone number (see directories in test_data_dir/). OTP is not validated.
• On successful login, the session is stored in memory for the duration of the server run.

Running the Server

Prerequisites

• Go 1.23 or later (installation instructions)

Install dependencies

go mod tidy

Start the server

FI_MCP_PORT=8080 go run .

The server will start on http://localhost:8080.

Usage

• Follow instructions in this guide to setup client
• Replace url with locally running server, for example: http://localhost:8080/mcp/stream
• When prompted for login, use one of the above phone numbers
• Otp/Passcode can be anything on the webpage

Simple curl client

curl -X POST -H "Content-Type: application/json" -H "Mcp-Session-Id: mcp-session-594e48ea-fea1-40ef-8c52-7552dd9272af" -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"fetch_bank_transactions","arguments":{}}}' http://localhost:8080/mcp/stream

If you run it once you will get login_url in response, running it again after login will give you the data

Simple fastmcp python client

from mcp.client.streamable_http import streamablehttp_client
from mcp.client.session import ClientSession
import asyncio
 
async def main():
    try:
        async with streamablehttp_client("http://localhost:8080/mcp/stream") as (
            read_stream,
            write_stream,
            _,
        ):
            async with ClientSession(
                read_stream,
                write_stream,
            ) as session:
                await session.initialize()
                tools = await session.list_tools()
                print(tools)
    except Exception as e:
        print(f"error: {e}")

if __name__ == "__main__":
    asyncio.run(main())

Requirements:

• fastmcp (pip install mcp)

Simple google adk example

from google.adk.agents.llm_agent import LlmAgent
from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset, StreamableHTTPConnectionParams

toolset = MCPToolset(
            connection_params=StreamableHTTPConnectionParams(
            url = "http://localhost:8080/mcp/stream"
            )
        )
root_agent = LlmAgent(
    model='gemini-2.0-flash',
    name='fi_mcp',
    instruction="You are an agent proficient in financial advises",
    tools=[toolset],
)

Requirements:

• google-adk(>=1.8.0) (pip install google-adk)
• export GOOGLE_API_KEY=YOUR_API_KEY
• Run adk web in parent directory

If you are using skip_auth branch you should directly get data as shown

FAQ

• Why do I need to login everytime in ADKs?

  A session in MCP is one to one connection between MCP server and MCP client and needs login once. If your ADK have multiple agents and you are creating multiple clients for them make sure you maintain a common sessionId and pass that around. But in case you want to skip auth entirely, although not recommeneded because we want your agents to work on production usecases, you can refer to this issue on how to skip auth

• Why am I getting invalid session id?

  If you are creating custom session id before client initialization then make sure you prefix it with mcp-session-. For example: 3ef38b37-323a-4bbd-acbb-3fe02f97783f is not valid and mcp-session-3ef38b37-323a-4bbd-acbb-3fe02f97783f is valid