Firewire Firewire
Help Center/ Getting Started & Features/ MCP Tools

Model Context Protocol (MCP) & Financial AI Tools

Complete architectural guide to Model Context Protocol (MCP) within Firewire: how generative neural models invoke deterministic backend functions, public portfolio analysis methods, and BYOK LLM key integration.

Section: Getting Started & Features · Updated October 2, 2026
Summary (TL;DR):

Model Context Protocol (MCP) is an open specification enabling AI Copilot to execute verified financial computation modules in Firewire as deterministic tool calls. Instead of hallucinating return percentages or bond durations, LLMs retrieve exact mathematics from Firewire's verified backend engine.

1. How MCP Operates in Firewire Architecture

Large language models excel at synthesizing complex findings and articulating financial rationale, yet struggle with precision math. Firewire solves this through a ReAct (Reasoning + Acting) agent framework coupled with MCP:

  • Reasoning: the agent parses the user's intent to identify required metrics (e.g. HHI concentration or XIRR).
  • Action (MCP Tool Call): Copilot constructs a structured tool invocation with specific parameters directed to Firewire's MCP server.
  • Execution: the platform calculates results deterministically using real-time market feeds and vetted quantitative finance formulas.
  • Observation & Response: structured outputs return to the model, producing clean, verifiable analytical answers.
Architecture of MCP tool invocations between Copilot agent and Firewire deterministic backend

2. Public MCP Tools for Sharing and Comparison

Version 1.75.27 introduces specialized tools designed for publicly accessible portfolio data:

MCP Tool Parameters Returned Analytical Intelligence
get_shared_portfolio_analysis share_token (access token) Aggregated public portfolio allocation by asset class (stocks, bonds, funds), top weight percentages, HHI concentration, and weighted bond duration.
get_shared_portfolios_comparison share_tokens (array up to 5 tokens) Comparative matrix of asset allocations, TWR performance, volatility, risk scores, and overlapping holdings across multiple shared accounts.

3. Security and Permission Boundaries

Public MCP endpoints adhere to strict data privacy principles:

  • Private account isolation: public tools only inspect portfolios with explicitly generated public share tokens created in «Portfolio Sharing». All other user accounts remain inaccessible.
  • Zero absolute monetary leakage: under anonymous sharing mode, MCP responses return strictly normalized weights ($\sum w_i = 100\%$) and never reveal absolute currency values.
  • No historical cash flow exposure: public tools do not provide transaction logs, deposit dates, or private withdrawal histories.

4. Bring Your Own Key (BYOK) Setup

Connect your personal GigaChat account to unlock unlimited AI queries:

  1. Navigate to «Settings» → «MCP & AI Agents» in the web application.
  2. Provide your Auth Key or Client ID and Client Secret credentials.
  3. Select your preferred model: standard, GigaChat-Pro, or GigaChat-Max for heavy portfolio simulations.
  4. Once saved, all tool executions draw directly from your personal API quota without platform limits.
«MCP & AI Agents» settings view with API key authorization and model selector