Skip to main content

Programmable Distributed Hyperscale UTAP* for all your AI apps.

One copy for your every workload. Transactions, real-time analytics and your AI agent on the same live data, no ETL, no replica lag, no second system. ScramDB JIT-compiles SQL to native machine code with GPU execution built in, and speaks the PostgreSQL wire protocol, so your drivers, tools, and queries just work.

True HTAPZero ETLBranching + PITRJIT native codeGPU executionPostgres compatible
What ScramDB is about

One engine. Every job you were running five systems for.

Five things you get on day one, and how the engine delivers each.

01 · AI agents

Let AI Agents (literally) program your database

ScramDB is more than an agentic database. It is built with AI agents kept in mind. It has a literally direct interface for any AI agent to interact with and manipulate your data. Copy the prompt and let your AI agent start developing.

  • Direct agent interface
  • Copy the prompt
  • Programmable
  • Any PG driver
01 · AI agents01/05
PG wireone copyruns your codeAI agentany frameworkYour apppsql / JDBCSQL + UDFScramDBone endpointone live copy

An agent connects like any Postgres client, then reads and writes your data directly.

02 · Real-time analytics

Real-time Analytical Dashboards

Everything is snappy, instant, and always correct. While your applications and AI agents provide and generate data in your ScramDB environment, you see everything instantly. Create drilldowns and power your analytics on the very same live rows your writes just landed in.

  • Instant
  • Always correct
  • Drilldowns
  • One live copy
02 · Real-time analytics02/05
commitscanlive resultsWritesOLTPGPUvectorizedBI dashboardalways currentone live copyCPU coresno ETL, no second copy

Analytical queries run across every CPU core and the GPU, on your live data.

03 · Durability

Extremely durable & Disaster survivor

Power outages, crashes, and anything else your infrastructure can throw at it: ScramDB comes back with every committed transaction intact, and it puts itself back together without you on a call. Silent corruption is caught rather than served. And when the damage is something you did, point-in-time recovery takes you back to the moment before it went wrong, or branches the database as it stood back then so you can look without touching production.

  • Survives power loss
  • Self-healing recovery
  • Point-in-time recovery
  • Instant branching
03 · Durability03/05
timerestarts, replays the logrewind to any momentfork itbranchinspect it, production untouchedevery commit is durablecrashnothing lostsilent corruption is caught, not served

Committed means committed. A crash rebuilds itself, and you can rewind to any moment before it went wrong.

04 · Deploy anywhere

Easy to manage & deploy

Autodiscovering nodes, cross-zone deployable, and tolerant to failures. You can literally have a single cluster spanning the whole world, and ScramDB can still serve transactions.

  • Autodiscovering nodes
  • Cross-zone
  • Failure tolerant
  • Global cluster
04 · Deploy anywhere04/05
protocoleu-centralservingus-eastservingap-southdownone logical databaseautodiscovery, no seed list to maintaina zone drops, the cluster keeps committing

Nodes discover each other across zones, and the cluster keeps serving when one drops.

05 · Autonomous

Autonomous Database

You know how hard databases are to operate: disk backups, tuning database-specific values, transaction timeouts, and everything else. ScramDB is a self-driving database. Its Ejder™ system drives it for you, optimizing resource costs and letting your database keep everything operational.

  • Self-driving
  • Ejder™
  • Optimizes resource cost
  • Stays operational
05 · Autonomous05/05
decidekept or rolled backobservelive workloadretunememory / workersmeasureevery changeapplywhile servingEjder™self-drivingno tuning tickets, no 3am page

The engine reads its own workload and retunes itself as it runs, with no DBA in the loop.

01 · AI agents01/05
PG wireone copyruns your codeAI agentany frameworkYour apppsql / JDBCSQL + UDFScramDBone endpointone live copy

An agent connects like any Postgres client, then reads and writes your data directly.

UTAP - Unified Transactional Analytical Processing

One database to rule them all.

Snowflake and Databricks spent a combined $1.25B concluding you need two systems plus a pipeline. ScramDB keeps the full PostgreSQL contract and does both jobs on one copy of your data, JIT-compiled, GPU-accelerated, built for AI-scale.

UTAP

Transactions, Analytics, AI. One Copy.

Transactional writes, analytical scans, and AI/vector queries all hit one live copy at the same instant, no CDC pipeline, no replica, no export, no lag. This is UTAP: Unified Transactional Analytical Processing.

TransactionsAnalyticsAIOne CopyRealtime
Full Postgres surface

Not Just the Wire, the Whole Contract

RBAC, row-level security, real triggers, stored procedures, MVCC isolation, and 2PC, the surface ClickHouse and Exasol drop and Cockroach only partly keeps. Connect with psql, JDBC, or any PG driver.

RBAC + RLSTriggers + ProceduresAny PG Driver
Proven

TPC-H, TPC-DS, TPC-C, All Passing

The full analytical and transactional suites, TPC-H, TPC-DS, and TPC-C, run correct and complete at every scale factor tested, SF1, SF10, and SF100, validated end to end, not a hand-picked subset.

H · DS · CSF1 · SF10 · SF100Correctness-Checked
1001010011010111000110101100
Compiled to the metal

JIT Native Code, Adaptive Tiers

Queries compile to native machine code in the background and switch tier per query, no config, cached to disk and reused across restarts. Interpreter cold, native hot, the engine decides.

Native CodeAdaptive JITDisk-Cached
Storage

Tundra: Hot, Warm, Cold, One Engine

The Tundra columnar engine tiers data by temperature automatically, hot rows in memory, warm on fast local storage, cold compressed out to cheap object storage, so you pay for capacity where it is cold and for speed only where it is hot.

Auto-TieringCloud TieringSkips Cold Blocks
GPU + parallelism

Every Core, and the GPU

Morsel-driven pipelines pin workers to CPU cores for linear scaling, and heavy batch work offloads to the GPU, baked into the engine across NVIDIA, AMD, and Apple Metal.

Core-PinnedLinear ScalingGPU Offload
AI-scale

Vectors + Agents, Native

ANN vector search rides the same MVCC/columnar write path as everything else, no bolt-on vector store. Built for agent-driven load: fast provisioning, RBAC as agent identity, an MCP interface on the roadmap.

Native ANNAgent-ReadyMCP
Storage IO

Built for NVMe, IO-Optimized

The storage path is tuned for modern NVMe: many parallel IO queues kept deep and saturated, so reads and writes stream at device speed instead of stalling on one lane.

Parallel QueuesDevice-SpeedDeep IO
ƒ(x)ƒ(x)ƒ(x)
Programmable

Sandboxed, Deterministic UDFs

Ship a UDF in the language you already write, running in a capability-sealed, deterministic sandbox, the real answer to Postgres's untrusted-C-extension footgun. A package declares what it may touch, and gets nothing else.

SandboxedDeterministicEight Languages
Batteries included

Search, vectors, and ML ship in the box.

Fifteen package families ship with the engine as capability-sealed, deterministic UDFs you call straight from SQL. Full-text, vector similarity, geospatial, clustering, regression, and streaming sketches, no extensions to compile, no sidecar services.

Search

Tokenizing + BM25 relevance ranking

ftsbm25
Vector / AI

Cosine + dot over embeddings

embeddings
Geospatial

Haversine distance + spatial cells

geoh3
ML + stats

Clustering, regression + anomaly scoring

kmeansdbscanregressionanomaly
Streaming sketches

Cardinality, quantiles, heavy hitters

hyperloglogthetatop-kquantile
Data hygiene

Masking, hashing + schema validation

anonymizejsonschema