AI Tool Discovery

Find Every AI Tool on Every Endpoint - Known or Unknown

Five independent detection layers find AI coding tools by name, by network destination, and by behavior - so a tool is caught even when no signature exists for it.

Book a Demo Start Free Trial

The Problem

Shadow AI Is the Default State

Developers adopt AI tools the way they adopt editors: individually, immediately, and without asking. A single engineering org routinely runs Claude Code, Cursor, Copilot, Windsurf, and a handful of local inference servers at the same time - none of it in an inventory, none of it in a risk register.

Signature-based inventory alone cannot close this. New AI tools ship weekly, and a fifteen-line Python script that calls an LLM API directly has no signature at all. Detection has to work by name, by destination, and by behavior simultaneously.

A tool nobody knows about cannot be governed, budgeted, or included in an audit. Every downstream control - DLP, policy enforcement, cost attribution, compliance evidence - depends on discovery being complete first.

How It Works

Five Independent Detection Layers

Each layer works on its own. Together they close the gap between the tools you approved and the tools that are actually running.

Discovery	Filesystem scanning for config directories, VS Code and JetBrains extensions, and process names - 48 AI tools detected by name
Network	Connection monitoring identifies 115 AI service endpoints across 69 providers, plus 7 local inference servers, by destination
Behavioral	Behavioral heuristics recognize how AI tools act - catching unknown and custom tools no signature list has heard of
Telemetry	Deep session telemetry for Claude Code - full conversation replay, token tracking, MCP visibility
Dependency	Dependency monitoring across 9 package formats, with known vulnerabilities flagged automatically

Detection depth varies by tool

Claude Code gets full session reconstruction: every prompt, response, tool call, file access, MCP connection, token count, cost, and subagent relationship. Codex CLI and Aider sessions are captured too. Other tools are detected by process, extension, and configuration directory - presence, version, and usage rather than conversation content.

Windows Subsystem for Linux is covered with full cross-boundary detection: Claude Code sessions running inside WSL are captured with the same depth as native sessions.

What You See

Fleet-Wide AI Inventory

Per-device AI tool inventory listing each detected tool with version, first seen, and last seen dates — Per-device AI tool inventory - every tool found by name or by network destination, with version and first and last seen.

AI system register listing discovered AI systems with risk tier, NIST impact, approval status, and device count — AI system register - every discovered AI system with risk tier, NIST impact, approval status, and how it was found.

Technical Specification

What Is Detected

Detection depth for selected tools

Detection depth for selected tools
Tool	Detection depth
Claude Code	Full session reconstruction - prompts, responses, tool calls, file access, MCP connections, token counts, cost, subagent relationships
GitHub Copilot	Process detection, VS Code extension monitoring, usage tracking
Cursor	Process detection, configuration directory monitoring
Aider	Process detection, config file monitoring, session capture
Windsurf / Codeium	Process and VS Code extension detection
Continue	VS Code extension and configuration monitoring
Cline	VS Code extension detection
Amazon Q Developer	AWS configuration and process detection
Codex CLI (OpenAI)	Process detection, session capture
Sourcegraph Cody	VS Code extension detection
Supermaven, Tabnine, Tabby	Process and extension detection
OpenHands	Process and Docker detection

Coverage

Tools by name (48): Claude Code, Codex (OpenAI CLI), Aider, OpenCode, OpenHands, Devin, Factory Droid, Amp (Sourcegraph), Goose (Block), Crush (Charm), Qwen Code, Kilo Code, Roo Code, Cline, OpenClaw, Kiro, Google Antigravity, Warp, Cursor, Windsurf, Codeium, GitHub Copilot, Cody (Sourcegraph), Continue, Tabby, Tabnine, Supermaven, Amazon Q Developer, Gemini Code Assist, Zed, Augment Code, Blackbox AI, Pieces for Developers, Trae, Qodo Gen, JetBrains AI Assistant, JetBrains Junie, Claude Desktop, ChatGPT Desktop, Microsoft Copilot, Grok (xAI), Ollama, LM Studio, Jan.ai, GPT4All, Msty, AnythingLLM, Open WebUI
Local inference servers (7): Ollama, LM Studio, Jan.ai, GPT4All, llama.cpp, text-generation-webui, vLLM
AI service endpoints (115, across 69 providers): OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Groq, Mistral, Cohere, DeepSeek, Perplexity, Together, Fireworks, Replicate, HuggingFace, OpenRouter, xAI, and more
Local model metadata: Model name, size, quantization, and last-used time for self-hosted inference
Dependency formats: 9 package formats with automatic known-vulnerability checks
WSL: Full cross-boundary detection

Compliance Mapping

Controls This Evidences

ISO A.5.9: Inventory of information and other associated assets
SOC 2 CC6.1: Logical access controls - which AI tools are permitted where
EU AI Act: AI Register auto-population from fleet discovery, with review queue

Session IntelligenceWhat the AI actually did Data EgressWhere the data went Shadow AI use caseThe full discovery workflow

See Kraitos AIDR in Action

Enroll an endpoint and see which AI tools it runs. Stop guessing what your AI-augmented organization is doing.

Book a Demo Start Free Trial

[email protected]kraitos.io