SPEC.md — Agent Police Department (Chrome Extension)

Product Identity

Name: Agent Police Department
Tagline: A police dashboard for your team's AI coding agents
Slug: agent-police-department-claude-code-agent-governan
Form: Chrome Extension (Manifest V3)
Tech Tier: Simple

Problem Statement

Teams adopting AI coding agents (Claude Code, Cursor, GitHub Copilot, etc.) lack visibility and governance. 65% of enterprises had security incidents from AI agents by April 2026. Only 30% have comprehensive governance. Small teams (5–50 devs) are priced out of enterprise tools ($10+/seat) and find open-source CLI tools too hard to use.

Solution

A lightweight Chrome extension that provides real-time visibility into AI coding agent activity, audit trails, and policy violation alerts — all running locally in the browser without a backend server.

MVP Scope (v1.0)

Core Features

Agent Activity Monitor
- Content scripts detect AI agent interactions on Claude.ai, GitHub, and ChatGPT
- Captures: tool use, file access, API calls, code generation events
- Timestamps every event with URL context
Policy Engine (6 Detectors)
- Permission Bypass: agent accesses files outside allowed scope
- Out-of-Scope: agent performs tasks beyond its mandate
- Self-Permissioning: agent attempts to modify its own permissions
- Disallowed Tools: agent uses blacklisted tools/APIs
- Redundant Operations: repeated identical actions
- Off-Task: agent drifts from assigned task
Real-Time Dashboard (Popup)
- Active session count and status
- Per-agent activity feed (last 50 events)
- Violation alerts with severity (Low / Medium / High / Critical)
- Team risk score (0–100)
- Compliance status indicator
Audit Log
- Local storage with 7-day retention
- Searchable by agent, URL, severity, detector type
- Export to JSON/CSV
- Hash-chained integrity verification (tamper-evident)
Policy Management
- Default policy YAML loaded on install
- User can edit rules via popup UI
- Rule types: allowed domains, allowed tools, file patterns, rate limits

Architecture

manifest.json (V3)
├── background.js (service worker)
│   ├── Event collector
│   ├── Policy engine
│   ├── Audit log storage
│   └── Alert dispatcher
├── content_script.js
│   ├── Claude.ai detector
│   ├── GitHub detector
│   └── ChatGPT detector
├── popup.html + popup.js + popup.css
│   ├── Dashboard UI
│   ├── Policy editor
│   └── Audit log viewer
├── options.html + options.js
│   └── Advanced settings
├── lib/
│   ├── detectors.js (6 violation detectors)
│   ├── policy.js (rule engine)
│   ├── audit.js (storage + integrity)
│   └── hash.js (sha256 chain)
├── default_policy.yaml
└── tests/
    ├── detectors.test.js
    ├── policy.test.js
    └── audit.test.js

Tech Stack

Chrome Extension Manifest V3
Vanilla JavaScript (no build step required for MVP)
Chrome Storage API (local)
Web Crypto API (SHA-256 for integrity)
Jest for unit tests
ESLint for linting

Data Model

{
  "event": {
    "id": "uuid",
    "timestamp": "ISO8601",
    "agent": "claude-code|cursor|copilot|chatgpt",
    "url": "https://claude.ai/chat/...",
    "type": "tool_use|file_access|code_generation|api_call|permission_request",
    "tool": "read_file|edit_file|bash|api_request",
    "target": "file_path_or_endpoint",
    "outcome": "allowed|denied|violation",
    "severity": "info|low|medium|high|critical"
  },
  "violation": {
    "id": "uuid",
    "event_id": "uuid",
    "detector": "permission_bypass|out_of_scope|...",
    "rule_triggered": "rule_id",
    "severity": "low|medium|high|critical",
    "message": "human readable explanation",
    "timestamp": "ISO8601"
  },
  "policy": {
    "version": 1,
    "rules": [
      {
        "id": "rule_1",
        "type": "allowed_domains",
        "enabled": true,
        "params": { "domains": ["github.com", "claude.ai", "api.openai.com"] }
      },
      {
        "id": "rule_2",
        "type": "disallowed_tools",
        "enabled": true,
        "params": { "tools": ["eval", "exec", "shell"] }
      }
    ]
  }
}

Build Requirements

npm test must pass (Jest)
npm run lint must pass (ESLint)
Anti-skeleton: every function has real implementation
No placeholder returns, no TODOs in shipped code
Secret scan: no API keys, credentials, or tokens in source

Distribution

Chrome Web Store (free tier)
GitHub release with built .zip
Source code at https://git.bunbunlabs.com/bunbun/agent-police-department-claude-code-agent-governan

Future Roadmap (post-MVP)

Team sync via cloud backend
Slack/Discord alerts
GitHub PR annotations
SSO integration
SIEM export (Splunk/Datadog)
Mobile push notifications

Verdict

Buildable: Yes — Chrome extension, no backend required
Gated: No — local-only for MVP, no server/DB/auth needed
Category: Chrome Extension

5.1 KiB Raw Blame History Unescape Escape