Documentation Architecture News & Updates Pricing About Speak to an engineer

Now in Early Access

The ultra high concurrency execution layer for AI agents

Give every agent an isolated environment to search, query, join, and analyze data without waiting on shared compute. Run SQL, vector, full-text, and geospatial workloads across your existing data platform.

See benchmarks

Compute moves to the agent

Async by default

Dispatch any query, load, or index build as a background job and get notified when complete.

Isolated per agent and session

Every agent gets its own scope. No shared state, no contention.

High-concurrency execution

Thousands of sessions run in parallel, with no warehouse queue.

Isolated execution,
governed access

Isolated by default

Each session runs in its own environment with read-only access to your systems of record.

Governed and auditable

Every action is recorded, and you can deploy Hotdata in your own cloud.

Unified execution

SQL, vector, full-text, geospatial, and cross-source joins in a single query.

Architecture Diagram

Fully managed or deployed in your own cloud environment.

Built with Rust, Apache Arrow, and Apache DataFusion

Keep your data platform. Add agent-native execution.

Hotdata runs on top of the platforms, catalogs, and formats you already use.

Compatible with

Snowflake

Databricks

PostgreSQL

MySQL

Neon

Supabase

Kafka

MQTT

AWS Glue

Unity Catalog

Polaris

DuckLake

Apache Iceberg

Delta Lake

CSV

Parquet

Start executing in one command

Install the CLI or hand the skill to your agent, then load data, index it, and run search and SQL against it, with no infrastructure to stand up.

Install with Homebrew

brew install hotdata-dev/tap/cli

Load SF Airbnb listings, search, query

hotdata auth register

hotdata databases create \
    --catalog airbnb

hotdata databases load \
    --catalog airbnb \
    --table listings \
    --url https://hotdata.dev/data/sf-airbnb-listings.parquet

hotdata indexes create \
    --catalog airbnb \
    --table listings \
    --column description \
    --type bm25

hotdata search "cozy apartment with a view" \
    --table airbnb.public.listings \
    --column description \
    --select name,description

hotdata query \
    "SELECT neighbourhood_cleansed AS neighbourhood, COUNT(*) AS listings, ROUND(AVG(review_scores_rating), 2) AS avg_rating FROM airbnb.public.listings GROUP BY neighbourhood_cleansed ORDER BY listings DESC LIMIT 5"

Pay for execution,
not idle capacity.

Storage

Object storage for managed databases and files.

10 GB

$0.20/mo

1 GB10 GB100 GB1 TB10 TB100 TB1 PB

Reads

Bytes scanned per query, summed across all queries.

1 TB

$10.24/mo

1 GB10 GB100 GB1 TB10 TB100 TB1 PB

Cost Breakdown

Base platform32%

Storage1%

Reads66%

Monthly Estimate

$15.44

/month

Base platform

$5.00

Storage10 GB

$0.20

Reads1 TB

$10.24

TotalFREE

What builders are saying

The combination of developer experience and performance stood out to us. The APIs were simple, query performance exceeded our expectations, and our team was able to integrate and start building immediately.

Elvis KahoroDeveloper Experience and Ecosystem Lead · dltHub

Questions?

Agents don't run one-and-done queries. They explore, branch, retry, and build intermediate context, often thousands of sessions concurrently. On shared warehouse compute, that work queues behind other jobs and human analysts. A dedicated execution layer gives each agent its own isolated, ephemeral environment so exploration is fast and one agent's workload never contends with another's.

Hotdata is built in Rust on Apache DataFusion and Apache Arrow, with Arrow-native execution over Parquet and open lakehouse formats. Execution environments start in milliseconds instead of waiting on a warehouse to warm up, and results are cached and reusable within a session. Our published benchmarks show flat latency under concurrency (holding ~2.7s while Snowflake climbs to 9.26s at 50 concurrent streams) and sub-100ms filtered vector search.

An Agent Session is the primary way agents work with data on Hotdata, the unit of isolation, execution, and teardown. Within a session an agent gets its own environment to run SQL, vector, full-text, and geospatial queries, create temporary tables and cached results, and join across sources. Instant databases, caching, and scratch tables are underlying capabilities the session uses. When the task is complete, the environment is torn down.

No. Hotdata is complementary. Your warehouse, lakehouse, catalogs, and pipelines stay in place and remain the source of truth. Snowflake and Databricks handle human analytics and governance; Hotdata adds an ultra high concurrency, isolated execution layer for AI agents that reads from what you already run.

Not necessarily. Hotdata reads from your existing sources, and attached connections stay live, queried in place rather than copied. When it helps, an agent can also load files like Parquet directly into a session for fast scratch work. Either way, your systems of record stay authoritative.

Each Agent Session runs in its own environment, so sessions can't see or interfere with one another. Access to your connected sources is read-only against snapshots, so agents can't change, remove, or overwrite data at the source. Every action is recorded in an audit trail.

Inexpensive object storage is billed separately from ephemeral compute, and compute only runs while a session is working. For bursty agent workloads (short spikes of heavy exploration rather than a steady load), you're not paying to keep a warehouse warm between tasks, which can meaningfully lower cost. Savings depend on your workload shape; the calculator on the pricing section lets you estimate yours.

Hotdata is built for Parquet and DuckLake, and works with Snowflake, Apache Iceberg, Delta Lake, AWS Glue, Unity Catalog, Polaris, and Apache Kafka. Ingestion pipelines including dlt, Fivetran, Airbyte, and Debezium land data into the formats and catalogs Hotdata reads.

Yes. With bring-your-own-cloud we deploy into your environment and operate the service, and you keep control of infrastructure, access, and keys. If you'd rather not manage infra, we also offer a fully managed option.

Speak to an engineer