Content
# AuditCopilot: Internal Control Insight Generator (MCP + A2A)
## 1. Purpose
**AuditCopilot** is a lightweight educational prototype that demonstrates how autonomous AI agents can exchange structured audit evidence and generate findings using two emerging interoperability ideas:
- **MCP (Model Context Protocol)** — a standardized interface for exposing tool capabilities to AI agents.
- **A2A (Agent-to-Agent) Messaging** — a pattern for structured communication between cooperating agents.
The goal is to illustrate how an auditor-style agent could gather system evidence, analyze risks, and output a concise report.
---
## 2. High-Level Architecture
Below is the system diagram that summarizes how the main components interact:
---
## 3. Key Concepts
### Model Context Protocol (MCP)
Modern AI systems like Large Language Models (LLMs) are powerful, but they have one big limitation — they mostly know what they learned during training.
They don’t automatically have access to **live data**, **company systems**, or **specialized tools**.
This means that, on their own, they can’t always give the most accurate or up-to-date answers.
The **Model Context Protocol (MCP)** solves this problem by creating a **standard way for AI models to connect with real-world data and tools**.
With MCP, an AI doesn’t just “guess” or rely on memory — it can directly fetch information, run checks, or use APIs that it’s allowed to access.
This makes AI systems **more relevant, trustworthy, and useful** in practical settings.
> *Reference: [Hugging Face MCP Course – Unit 1: Introduction](https://huggingface.co/learn/mcp-course/unit1/introduction).*
**Technical view:**
MCP defines a **tool interface** that lets AI agents safely call structured APIs.
In this demo:
- `MCPControlsServer` acts as a **mock MCP provider**.
- It exposes methods such as:
- `get_system_list`
- `get_control_state`
- `get_audit_log`
- These simulate system metadata, control states, and audit logs from external environments.
This helps the **EvidenceAgent** interact with “systems” in a reproducible, schema-driven way.
### Agent-to-Agent (A2A) Messaging
**Technical view:**
A2A defines how multiple agents exchange messages via a shared bus.
In this prototype:
- The **A2A Bus** is a simple in-memory message queue.
- **EvidenceAgent** fetches and packages evidence bundles.
- **AuditAgent** consumes those bundles, analyzes risks, and generates recommendations.
- The message schema (`evidence_bundle`, `findings`, etc.) mimics how independent agent processes could coordinate in a distributed system.
## 4. Components
| Component | Role | Notes |
|------------|------|-------|
| **A2A Bus** | Message transport between agents | Handles serialized audit messages |
| **EvidenceAgent** | Gathers evidence via MCP tools | Produces structured JSON evidence |
| **AuditAgent** | Consumes evidence and performs rule-based analysis | Writes rich table + Markdown report |
| **MCPControlsServer (mock)** | Simulated MCP tool endpoint | Supplies system data & control states |
| **sample_data/systems.json** | Static system metadata | Acts as mock evidence source |
| **run_demo.py** | Orchestrator script | Launches both agents locally |
---
## 5. Execution Flow
1. The user (auditor) runs:
```bash
python run_demo.py
2. The A2A Bus is initialized.
3. EvidenceAgent requests data from the mock MCP server.
4. AuditAgent consumes the evidence and identifies issues.
5. Both agents communicate through the A2A Bus.
6. The system prints a summary table to the console (using rich) and saves a report to:
reports/audit_report.md
## 6. Run demo:
```bash
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
python run_demo.py
```
## 7. Example Output
Console view:
| System | Env | Owner | Issues |
|----------------------|---------|---------------|-----------------------------------------|
| PaymentsService | prod | payments-team | TLS disabled, No authentication on APIs |
| CatalogService | staging | catalog-team | Backups disabled |
| NotificationService | prod | growth-team | No encryption at rest |
