ATLAS Governance Gateway (Phase 2)

A Secure Middleware for AI Agents, powered by FastMCP, Inngest, and Modal.

The ATLAS Governance Gateway is a centralized hub that intercepts AI Agent tool calls, evaluates potential risks using a generic LLM/SLM (hosted on Modal.com), and manages "Human-in-the-Loop" (HITL) escalations via durable Inngest workflows.

Logic Attribution: The governance logic and dataset design are attributed to Anna Ko anna_ko@berkeley.edu (UC Berkeley). This implementation strictly follows the "approve -> auto_approve" deterministic flow for low-risk actions per project v3.0 findings.

🚀 Key Features

• Durable Governance: Uses Inngest to manage long-running human review workflows that survive server restarts.
• FastMCP + FastAPI: Exposes Model Context Protocol (MCP) tools via a robust generic web server.
• Risk Evaluation: Real-time checking against EU AI Act standards via atlas-welfare-v1 on Modal.
• Compliance: Enforces Article 14 by pausing high-risk actions until human approval is received.
• Secure Deployment: Dockerized environment with secure credential management.

---

🛠️ Architecture

sequenceDiagram
    participant Agent as AI Agent (MCP Client)
    participant GW as Governance Gateway (FastAPI)
    participant Modal as Atlas Brain (Modal Evaluation)
    participant Inngest as Inngest Workflow
    participant Human as Sarah (Human-in-the-Loop)

    Agent->>GW: Call Tool (e.g. request_payment_extension)
    GW->>Inngest: Trigger Governance Event
    GW-->>Agent: Returns "PENDING REVIEW"
    
    Inngest->>Modal: Evaluate Risk (EU AI Act)
    Modal-->>Inngest: Risk Score (0-100)

    alt Risk Score < 70
        Inngest->>Inngest: Auto-Approve Action
    else Risk Score >= 70
        Inngest->>Inngest: PAUSE & Wait for Human
        Human->>GW: POST /webhook/approval (APPROVED)
        GW->>Inngest: Send Approval Event
        Inngest->>Inngest: Resume & Finalize
    end

The system consists of three main components:

1. Gateway (gateway.py): A FastAPI application that serves as the entry point.

   • MCP Server: Mounted at /mcp/sse to serve Agent tools (check_status, request_extension, etc.).
   • Inngest Endpoint: Served at /api/inngest to handle workflow triggers and step execution.
   • Webhook: Listens at /webhook/approval for external human approval signals.

2. Workflows (workflows.py): Defines the durable business logic.

   • handle_governance: The core workflow that:
     1. Calls Modal to assess risk.
     2. Auto-approves usage if Risk Score < 70.
     3. Pauses and waits for atlas/sarah.decision event if Risk Score >= 70.

3. Agent Integration:

   • Any MCP-compliant agent (Claude Desktop, cursor, or custom scripts) can connect to the Gateway.

---

📦 Installation & Setup

Prerequisites

• Docker Desktop installed and running.
• (Optional) python 3.10+ if running locally without Docker.

Quick Start (Docker)

We provide a setup.sh script to automate the configuration and deployment.

1. Run the Setup Script:

   ./setup.sh
   

2. Configuration: The script will prompt you for:

   • Modal Function Name: (Default: nislam-mics/ATLAS-NIST-Measure)
   • AI Backend: Choose OpenAI or Local LLM (Ollama).
   • Inngest Keys: Event Key and Signing Key (Required for Production).
   • Modal Credentials: Token ID and Secret.

   Note: The script creates a .env file automatically. DO NOT commit this file.

3. Usage: The script will build the Docker image atlas-gateway and run it on port 8000.

---

🔌 Endpoints

Once running, the Gateway exposes the following endpoints on http://localhost:8000:

Endpoint	Method	Description
/	GET	Health check and status.
/mcp/sse	GET	MCP Endpoint. Connect your AI Agent here.
/mcp/messages	POST	MCP Protocol messages (handled by FastMCP).
/api/inngest	POST	Inngest Webhook. Connects to Inngest Cloud.
/webhook/approval	POST	Human Approval Webhook. Simulate approval signals.

---

🧪 Testing the Governance Flow

1. Manual Verification

Check if the gateway is running:

curl http://localhost:8000/
# Expected: {"status": "ATLAS Governance Gateway Running", ...}

2. End-to-End Simulation

We included a script test_handshake.py that simulates the entire flow without needing an external Agent.

1. Ensure the Gateway is running (via Docker or locally).

2. Run the test script:

   python test_handshake.py
   

3. Flow:

   • Script simulates an Agent calling request_payment_extension (High Risk).
   • Gateway returns PENDING REVIEW.
   • Inngest triggers a workflow run (check Inngest Dashboard).
   • Script waits... then sends an Approval signal to /webhook/approval.
   • Workflow completes.

🤖 AI Agent Prompt Instructions

To ensure your AI Agent correctly uses the Governance Gateway, include the following instructions in the Agent's System Prompt:

🛑 Governance Protocol

1. This environment is governed by the ATLAS Governance Gateway.
2. When you call a payment or high-risk tool, the gateway may return a message starting with PENDING REVIEW (Ref: evt_...).
3. DO NOT assume the action failed.
4. DO inform the user that the request has been queued for human governance review and provide the Reference ID.
5. DO NOT retry the call immediately; wait for the system to process the background workflow.

---

🔄 Detailed Governance Procedure Call

The following sequence occurs when a high-risk tool is invoked.

1. Initial Tool Request (Agent -> Gateway)

Endpoint: SSE /mcp/sse (Streamed)

• Request (JSON-RPC):

  {
    "method": "tools/call",
    "params": {
      "name": "request_payment_extension",
      "arguments": { "beneficiary_id": "BEN-123", "reason": "Hardship" }
    }
  }
  

• Gateway Response:

  PENDING REVIEW (Ref: evt_48f1fb4b). Action queued subject to governance checks.
  

2. Risk Assessment (Inngest -> Modal)

Endpoint: Modal.com (Inference API)

• Input: Tool name and arguments.

• Assessment Result:

  { "risk_score": 85, "classification": "HIGH_RISK", "action": "PAUSE" }
  

3. Case A: HITL Required (Risk >= 70)

If the Atlas Brain determines the action is high-risk, a notification is sent to the human reviewer (Sarah).

• Sarah's Workflow:

  1. Sarah receives a notification via Inngest/Webhook.
  2. Sarah reviews the reason and beneficiary_id.
  3. Sarah approves: POST /webhook/approval

  { "decision": "APPROVED", "event_id": "evt_48f1fb4b" }
  

• Workflow Result: Workflow resumes and marks the tool execution as AUTHORIZED.

4. Case B: Auto-Approval (Risk < 70)

If the action is deemed low-risk (e.g., standard status check or well-documented low-value extension), the system skips human intervention.

• Workflow Result: The workflow proceeds immediately to completion without waiting for an external signal.
• Audit Log: "Auto-approved based on Risk Score [X] per Article 14 Compliance."

5. Workflow Resolution (Inngest)

The atlas/sarah.decision event resumes the workflow. The final authorization is logged in the Inngest dashboard.

---

📂 Project Structure

• gateway.py: Main server entry point (FastAPI + FastMCP).
• workflows.py: Inngest workflow definitions and client initialization.
• setup.sh: Automated setup and deployment script.
• Dockerfile: Container definition (Secure, no secrets baked in).
• requirements.txt: Python dependencies.
• test_handshake.py: Verification utility.

---

🔒 Security Notes

• Secrets: Never bake API keys into the Docker image. Always pass them via --env-file or environment variables at runtime.
• Inngest Signing: Production deployments MUST verify the INNGEST_SIGNING_KEY to prevent unauthorized workflow triggers. This is enforced in workflows.py.