Introduction

Visibility and control for every AI agent action in your organization

The accountability engine for AI Agents

AI agents are shipping to production with zero security telemetry. They execute processes, read and write files, make network calls — and nobody is watching. Quint changes that.

Quint is an endpoint agent that monitors AI coding assistants and autonomous agents on developer machines. It captures every process execution, file access, and tool invocation as structured events, tracks sessions across agent interactions, and surfaces risk through a centralized dashboard. Think of it as endpoint detection and response (EDR), purpose-built for AI agents instead of humans.

How It Works

End-to-end flow in one page — follow a single tool call from agent → interception → cloud scoring → dashboard.

Architecture

Three-tier design: endpoint agent (truth) + cloud API (aggregation) + dashboard (visibility).

Explore the Platform

Endpoint Security (ES)

Kernel-level ground truth on macOS: process spawns, file I/O, code-signing verification.

Network Extension (NE)

Transparent interception of LLM API traffic on macOS — no env vars, no CA trust dance.

Event Ingestion

How the daemon delivers events to the cloud — batching, overflow, SNS/SQS fan-out.

Edge Daemon

Unifies proxy + ES into one session model, forwards to cloud, holds the signed audit log.

Cloud Learning

Cross-session divergence detection and the fleet learning loop. The cloud sees only intent frames, never conversation content.

Dashboard

Session-centric visibility across every agent, machine, and alert in your org.

How It Works

Quint operates as a system daemon that monitors AI agent activity at the OS level:

The daemon detects AI agents through code signing identification — no proxy configuration, no code changes, no agent cooperation required. When an agent like Claude Code spawns a process or accesses a file, Quint captures it as a structured event with full context.

Three-Tier Architecture

Quint’s architecture separates concerns into three layers:

1. Endpoint Agent (Truth) — The daemon running on each machine. Captures ground-truth system events via Endpoint Security on macOS (Linux via eBPF is on the roadmap — see Platform Coverage). This is the source of truth for what actually happened.

2. Cloud API (Aggregation) — Go service at api.quintai.dev. Receives events and session lifecycle data, stores in PostgreSQL with row-level security, and serves the REST API for querying.

3. Dashboard (Visibility) — Next.js app at cloud.quintai.dev. Session-centric UX with drill-down from fleet overview to individual events.

Divergence between what an agent claims to do and what the OS observes is the key security signal.

Production Deployment Flow

Event Model

Every captured action is stored as an event with structured fields:

Field	Description
action_type	What happened: PROCESS_EXEC, FILE_READ, FILE_WRITE, FILE_DELETE, NETWORK_CONNECT, etc.
tool_name	The binary or tool that was invoked (e.g., git, node, curl)
arguments	Process arguments or file paths
session_id	UUID v5 linking the event to a session
agent_id	The detected AI agent that triggered the action
risk_score	Assessed risk level (0-100)

Tech Stack

Component	Technology
Endpoint Agent	Go daemon + Swift Endpoint Security & Network Extension (macOS). Linux (eBPF) + Windows (WFP/ETW) on roadmap.
Ingest	HTTPS batch POST to /v1/ingest → SNS FIFO → SQS fan-out
Cloud API	Go (stdlib net/http, pgx/v5)
Database	PostgreSQL (RDS) with Row-Level Security
Auth	Supabase (JWT) + Deploy Tokens (SHA-256 hashed)
Dashboard	Next.js on Vercel (shadcn/ui)
Config	/etc/quint/config.yaml

Risk Score Bands

Score	Risk Level	Color
0-10	None	Green
11-30	Low	Blue
31-55	Medium	Yellow
56-80	High	Orange
81-100	Critical	Red

The risk score is a shadow-mode corroborating signal, not the enforcement mechanism. Enforcement is deterministic scope evaluation — only an out-of-scope action against a sensitive resource is block-capable. See Intent vs Truth.