# Topic 29: Groq Tool Use & MCP Strategy

**Audit Date**: 2026-02-01
**Auditor**: Agent Antigravity
**Scope**: Tool Calling, MCP, Remote vs Local Execution

---

## 1. Tool Use Architecture
The Sentinel System implements **Local Tool Calling**, where the LLM defines the action, but the application code executes it. This provides maximum control and security for forensic operations.

### 1.1 Native Groq Tool Use (Local)
**Implementation**: `GroqClient.generate_tool_call`
**Compliance**: ✅ **Pass**.
The system correctly implements the chat completions tool format:
*   **Parallel Execution**: explicit support via `parallel_tool_calls` parameter.
*   **Retry Logic**: Smart handling of 400 errors with temperature scaling.
*   **Rate Limits**: Telemetry integration.

```python
# app/core/llm_client.py
async def generate_tool_call(self, ...):
    payload = {
        "tools": tools,
        "tool_choice": "auto",
        "parallel_tool_calls": True # Feature flagged
    }
    # ...
```

### 1.2 Remote MCP (Model Context Protocol)
**Implementation**: **Detected Gap**.
The `ModelRegistry` lists `Capability.REMOTE_MCP` for models like `llama-3.3-70b`, but the `GroqClient` logic currently **does not** support the specific `type: "mcp"` tool definition or the `responses.create` endpoint required for Server-Side MCP orchestration.

*   **Current State**: System uses Local Execution (Standard Function Calling).
*   **Missing**: Logic to construct `{"type": "mcp", "server_url": ...}` payloads.
*   **Impact**: No access to third-party MCP servers (Stripe, Firecrawl) without custom local connectors.

**Recommendation**:
For Phase 7 (Future Expansion), implement a dedicated `generate_mcp_response` method in `GroqClient` that leverages Groq's `responses.create` API to offload orchestration.

---

## 2. Groq Built-In Tools
**Implementation**: `GroqClient.generate` (via `compound_custom`)
**Compliance**: ✅ **Pass**.
The system supports Groq's unique `groq/compound` models by injecting `compound_custom` capabilities when `enabled_tools` are provided.

```python
# app/core/llm_client.py
if enabled_tools:
    payload["compound_custom"] = {
        "tools": {"enabled_tools": enabled_tools}
    }
headers["Groq-Model-Version"] = "latest"
```

---

## 3. Security & Control
Sentinel prioritizes **Local Tool Execution** for security reasons. By executing tools locally:
1.  **Data Sovereignty**: Sensitive PI I/scam data does not leave the secure enclave to remote MCP servers.
2.  **Forensic Logging**: Every tool execution is logged internally for chain-of-custody.
3.  **Approval Gates**: Critical actions (e.g., "Report Takedown") can be intercepted by human operators (Human-in-the-Loop), which Remote MCP "Auto-Approval" might bypass.

---

## 4. Conclusion
The system effectively uses **Local Tool Calling** for its internal forensic tools. While **Remote MCP** is not implemented, this is an acceptable architectural choice for a high-security application.

**Status**: **PASSED (with MCP Roadmap Note)**.