Find the AI agents on your machine, and see what they can touch. Enterprise-style AI agent visibility for freelancers, students, small businesses, and solo devs, free and local.
v1.0 scan. In v1.1 the summary reports installed agents and running processes separately.
$ python3 ai_guardian.py ai-guardian 1.0.0 - what AI is on this machine, and what can it touch? AI apps (3) • Claude Anthropic Claude desktop • LM Studio LM Studio local AI models • Ollama Ollama local AI models Running now (6) • Claude running now (26 processes) • LM Studio running now (4 processes) ... Local AI servers (2) • Ollama listening - port 11434 • LM Studio listening - port 1234 MEDIUM AI extension can read every website: Claude Fix: set Site access to 'On click' in Chrome Summary: 16 AI agent(s) found · 0 high, 1 medium risk(s)
16 AI agents on one personal laptop. Most people would guess two or three.
AI apps, command-line agents, running processes, AI browser extensions, local model servers, and MCP servers: the plugins that connect agents to your files and accounts.
Flags plain-text API keys, agents with access to your whole drive or SSH keys, permission prompts turned off, and AI servers open to your Wi-Fi.
Phone alerts when an AI process reads your keys or wallets, launches a shell, or talks to a server it never has before.
| Level | Risk | Why it matters |
|---|---|---|
| HIGH | API keys in plain text in agent configs | Any app or malware that reads your files can steal them |
| HIGH | Agent plugin with access to your whole drive or ~/.ssh | One poisoned webpage away from leaking your keys |
| HIGH | Permission prompts turned off | The agent runs commands without asking |
| HIGH | Local AI server open to your network | Anyone on coffee-shop Wi-Fi can use it |
| HIGH | Agent plugin over unencrypted http:// | Traffic can be read or changed in transit |
| MEDIUM | Plugins that auto-install the latest code | A hijacked package runs with your agent's access |
| MEDIUM | AI extensions that can read every website | Includes your email, bank, and work tools |
| LOW | Agent plugins that send data to an outside server | Your files and messages leave your computer |
git clone https://github.com/jeffreyesdavid/ai-guardian.git cd ai-guardian python3 ai_guardian.py # full scan python3 ai_guardian.py --json # machine-readable output python3 ai_guardian.py --fail-on medium # stricter exit code for CI
The day M8ven added ai-guardian to its MCP Trust Index, I installed M8ven's own ToolCheck server and scanned before and after. ai-guardian caught it at once, and the diff revealed a flaw in my own tool: the agent count dropped after I added one, because live processes shared a total with installed agents. v1.1 fixes that, flags plugins that send data to outside servers, and gives real version-pinning advice.
At Fal.Con 2026, CrowdStrike launched Falcon Guardian to help enterprises discover their AI agents, map what they can access, and stop threats where they run. The idea is right. But the people installing AI agents fastest (freelancers, small shops, students) will never buy enterprise security. ai-guardian brings the same discover → map → watch model to everyone else, for free.
| ✓ v1 | Discover + Map: find AI agents and flag risky setups |
| ✓ v1.1 | Accuracy: installed vs. running counts, outside-server detection, real version pinning |
| v2 | Watch: live phone alerts on sketchy AI behavior |
| v3 | Explain: a local AI describes each alert in plain English |
| v4 | Red vs. blue lab: agents attack each other with prompt injection to build detection rules |