Guarded Agent

Run Locally

Prerequisites

• Python 3.10+
• Node.js 18+

Step 1 — Install Python dependencies

cd Assignment
pip install -r requirements.txt

Step 2 — Install frontend dependencies

cd frontend
npm install

Step 3 — Start MCP Notes Server (Terminal 1)

cd Assignment
python -m backend.custom_mcp.notes_server

Runs at http://127.0.0.1:8002/mcp

Step 4 — Start Backend (Terminal 2)

cd Assignment
uvicorn backend.main:app --reload --host 0.0.0.0 --port 8000

Runs at http://localhost:8000

Step 5 — Start Frontend (Terminal 3)

cd frontend
npm run dev

Runs at http://localhost:5173

---

Testing the Agent

Type these in the Chat box on the dashboard:

Command	Tool Called
create note title: X content: Y	create_note
show all notes	get_notes
search milk	search_notes
update note 1 title: X content: Y	update_note
delete note 1	delete_note

---

Testing Policy Features

Block a tool

Invoke-RestMethod -Uri "http://localhost:8000/api/settings/" -Method POST -ContentType "application/json" -Body '{"key": "blocked_tools", "value": "delete_note"}'

Then type delete note 1 — will be blocked.

Unblock

Invoke-RestMethod -Uri "http://localhost:8000/api/settings/" -Method POST -ContentType "application/json" -Body '{"key": "blocked_tools", "value": ""}'

Enable human approval

Go to Settings in dashboard → set enable_human_approval = 1 → Save. Then try delete note 1 — will require approval from the Approvals panel.

Prompt injection detection

Type this in chat:

Will be blocked automatically with: Prompt injection detected in user message.

---

Policy Engine

The policy engine (policy_engine.py) enforces these rules in order:

1. Global toggle — enable_tool_access = 0 blocks all tools
2. Prompt injection detection — blocks known injection patterns
3. Argument size limit — blocks tool args over 2000 chars
4. Blocked tools list — blocked_tools = comma-separated tool names
5. Human approval — enable_human_approval = 1 requires approval for delete/update

Rules are checked on every request. Dashboard changes take effect immediately without restart.

---

Edge Cases

Scenario	Behavior
MCP server crashes mid-call	MCPClient worker reconnects automatically on next request
Prompt injection attempt	Detected by regex patterns, blocked before tool execution
Conflicting rules	blocked_tools takes priority over approval requirement
Approver offline	Request stays pending in DB indefinitely until approved/denied
Tool not recognized	Returns helpful message with supported commands