WhatPowerstheBrain.

DuckDB execution engine. Rust concurrency layer. Apache Iceberg storage. 13-step query pipeline. Here's how it all fits together.

DuckDB + Rust
13-Step Pipeline
Enterprise Security

Integrations

Works With Your Agent Framework

First-class support for the tools your agents already use.

LangChain
CrewAI
AutoGen
MCP
dbt
Python
TypeScript

QUERY PIPELINE

13 Steps. Every Query. Every Time.

Every SQL statement — from agent or human — passes through the same hardened pipeline. Security, governance, and billing are never optional.

01SemaphoreConcurrency gate
02Table ExtractParse referenced tables
03Policy CheckABAC enforcement
04Cost EstimatePre-execution costing
05Quota CheckRate + credit limits
06Row FilterRow-level security
07TranspileSnowflake → DuckDB SQL
08Snapshot PinIceberg snapshot isolation
09DuckDB ExecuteVectorized columnar engine
10AI HealAuto-fix failed queries
11Column MaskPII/sensitive redaction
12MeterPer-second billing
13AuditImmutable S3 log
Concurrency & Audit
Security & Governance
Billing & Quotas
SQL Intelligence
Execution & Storage

THE ENGINE ROOM

Built on Giants. Engineered for Production.

We didn't reinvent the wheel. We built the car. HatiData combines the raw vectorized speed of DuckDB with the concurrency, security, and state management required for the modern cloud.

RustDuckDBApache Iceberg
Powered by DuckDB

The Kernel: Vectorized Execution

We leverage DuckDB’s columnar, vectorized execution engine to process analytical queries at the speed of memory.

  • Zero-Copy: Data is processed in Apache Arrow format.
  • SIMD Optimized: Parallel instruction execution for aggregations.
  • MIT Licensed: Open standards. No proprietary lock-in.
Orchestrated by Rust

The Brain: Concurrency & Safety

DuckDB is single-process. HatiData makes it cloud-native. Our custom Rust proxy handles connection pooling, async scheduling, and fault tolerance.

  • Stateless: Nodes spin up/down in milliseconds.
  • Safe: Memory-safe concurrency prevents crashes under load.
  • Smart: Automatic query routing based on data locality.
Grounded in Iceberg

The State: Governance & Storage

We decouple compute from storage completely. Your data lives in object storage; we just borrow it.

  • IAM Native: We use cloud-native IAM (AWS, GCP, Azure), not static keys.
  • ACID Compliance: Full transactional integrity via Apache Iceberg.
  • Infinite Scale: Storage scales independently of compute.

Why HatiData?

DuckDB is an incredible engine. But production workloads need more than an engine — they need a platform.

FeatureRaw DuckDB (DIY)HatiData (Enterprise)
Execution EngineVectorized (Fast)Vectorized (Fast)
ConcurrencySingle-Process / LockedMulti-Node Auto-Scaling
Storage LayerLocal Files / Manual S3Managed Iceberg Catalog Sync
SecurityNone (File Permissions)RBAC, IAM, & SSO Integration
CachingOS Page CacheIntelligent NVMe Tiering
BillingN/APer-Second (No Minimums)

The “No Lock-In” Guarantee

We use the open-source DuckDB engine and the standard Apache Iceberg format — you are never locked into a proprietary ecosystem. You can read your HatiData tables with Spark, Trino, or a local Python script — anytime, anywhere.

SECURITY

Enterprise Security, From Day One

Not Phase 3. Not 'coming soon'. Day One.

CMEK

Your encryption keys. Your KMS. We never see them.

PrivateLink

Zero public internet traversal. Private connectivity only.

Immutable Audit

Every query logged. S3 Object Lock. 7-year retention.

RBAC + Masking

6 roles. Column-level masking. Row-level security.

DEPLOYMENT

One Command to Promote

Start local. Push to cloud or VPC with a single command. Data, schemas, and agent memories migrate seamlessly.

LocalDuckDB on your machine, MCP tools included
CloudManaged infra, shared memories, team dashboards
VPCYour network, your keys, your compliance
terminal
# Initialize locally
$ hati init
Created local engine at .hati/local.duckdb
# Agent memory works immediately
$ hatidata-mcp-server --local --agent-id my-agent
MCP server ready (24 tools, local mode)
# Push everything to cloud when ready
$ hati push --target cloud
Synced 3 tables + 1,247 memories to cloud

See it in action.

Install locally in 30 seconds. The full pipeline runs on your laptop.