System Architecture

Three-tier architecture: ES extension (truth), forward proxy (intent), and the cloud platform. Covers session detection, event ingestion, and the API service.

System Architecture

Quint is a three-tier security platform that combines OS-level ground truth (EndpointSecurity system extension), intent-level interception (forward proxy plus MCP gateway), and a cloud platform for fleet-wide visibility. The key security signal is divergence between what an agent claims to do and what actually happens on disk.

Target-state architecture

The diagram below is Quint’s enterprise-ready target state — the north-star topology the platform is converging on. Events flow from an MDM-managed endpoint fleet over the SNS/SQS bus into stream processing and tiered hot/warm/cold storage, with per-tenant egress to customer SIEMs and a dedicated fleet-management plane. Note that the behavioral stream ranks only and never makes an enforcement decision. The tiers described in the rest of this page are the current shipping implementation this target state builds on.

Quint — enterprise-ready target-state architecture.

Three-Tier Architecture

Tier 1: Edge

Runs on the user’s machine as a signed .pkg install.

• QuintAgent.app hosts two system extensions:
  • Endpoint Security for kernel-level process and file events
  • Network Extension for transparent LLM API interception
• Go daemon runs as a LaunchDaemon. It operates the forward proxy (MITM TLS plus MCP relay/gateway), the unified session tracker, and the cloud forwarder. Events stream to api.quintai.dev via HTTPS.
• Audit log at ~/.quint/quint.db with Ed25519 signatures and SHA-256 hash chaining. Every captured request, response, and tool call is attributable to a specific session.

Tier 2: Cloud Platform

Deployed on AWS ECS Fargate. Five services share one Postgres instance and one Redis cache:

Service	Role
ingest	Authenticate, validate, fan out to SNS
pipeline	Consume SQS, write actions table
session-processor	Upsert sessions table
alert-processor	Evaluate rules, write alerts
api	Dashboard and public API

Row-level security is enabled on all tenant tables using app.current_org_id. The actions table is partitioned by month. See Cloud Overview for the full model.

Tier 3: Dashboard

Next.js on Vercel. Reads from the API service over Supabase JWT or API key. Session-centric views with drill-down into each action. See Dashboard Architecture.

Signal: Intent vs. Truth

The proxy sees what the agent claims to do. The ES extension sees what the OS actually does. Divergence is the high-confidence signal.

Example: an agent that says “read config.json” but actually opens ~/.ssh/id_rsa. Tier 1 captures the tool call. Tier 2 captures the file open. The mismatch triggers an alert.

End-to-End Flow

See How It Works for a single tool call traced through all layers.

Deployment Topology

Production (AWS)

Service	Resources	Purpose
API Service	ECS Fargate	REST API, event ingest
Ingest	ECS Fargate	Token auth, validation, fan-out
Pipeline	ECS Fargate	SQS consumer, Postgres writer
Session Processor	ECS Fargate	Session lifecycle upserts
Alert Processor	ECS Fargate	Rule evaluation
PostgreSQL	RDS t4g.medium	Events, sessions, partitioned by month
Redis	ElastiCache t4g.micro	L1 cache, rate limiting, SSE pubsub
ALB	Elastic Load Balancer	HTTPS termination with ACM certificate

Infrastructure is fully Terraformed under quint-platform/infra/terraform/environments/prod/.

Endpoint (macOS)

Component	Footprint	Purpose
QuintAgent.app	~20MB disk	Container for system extensions
ES Extension	~50MB RAM steady	OS-level event capture
NE Extension	~40MB RAM steady	Transparent network interception
Go Daemon	~100MB RAM steady	Proxy, session tracking, forwarder
Local SQLite	10-100MB	Signed audit log (rolling)

Security Properties

Source code content, credentials, and full LLM conversation bodies never leave the machine. Only structured metadata (tool name, capability, risk decomposition) goes to the cloud.

• Audit log tampering: Ed25519 signature per row, chained by prev_hash. A single modification breaks the chain.
• TLS MITM: per-hostname leaf certs signed by a local CA that never leaves the machine.
• Cloud auth: deploy tokens are SHA-256 hashed in Redis, JWT for dashboard users.
• Tenant isolation: Postgres RLS on every tenant table, SNS/SQS FIFO ordering per session.

Design Principles

1. Local-first capture, cloud-first scoring. Raw bodies stay on the machine. Normalized events flow to the cloud.
2. Lossy at the edge, durable in the cloud. Under backpressure, the edge drops events rather than block user traffic. The cloud retries.
3. Zero client changes where possible. Forward proxy plus NE means no SDK integration, no new endpoints.
4. Session-centric model. Everything anchors to sessions, not raw events. Audit rows, cloud events, and dashboard views all join on session_id.
5. Code signing first. Agent detection uses macOS code signing as the highest-confidence signal, falling back to process name and path matching only when signing data is unavailable.
6. Tenant isolation at every layer. Deploy tokens are scoped to organizations. Postgres uses monthly partitions with row-level security. Cloud queues use FIFO ordering keyed by session.