DuckDB is a free, MIT-licensed analytical database that runs inside your application process rather than as a separate server. It is columnar and vectorised, so aggregations, joins and window functions over millions of rows run in milliseconds on a laptop, and it reads CSV, Excel, Parquet, JSON, Postgres, S3, Iceberg and DuckLake sources directly without an ETL step. The current stable release is 1.5.5 (July 2026); version 2.0 is due in autumn 2026. Amazon agreed to acquire DuckLabs, the company behind it, on 26 August 2026, with the project staying open source under the independent DuckDB Foundation.

Is DuckDB a data warehouse?

Not on its own. A warehouse is a shared, governed, multi-user service; DuckDB is an engine. It becomes warehouse-like when you add a serving layer (MotherDuck, the Quack client-server protocol, or a small API in front of it), a catalog (DuckLake) and a context layer for agents. For most businesses with a terabyte or less of data and a team you can name, that is enough. For petabytes, hundreds of concurrent users and row-level security, you still want Databricks, Snowflake or BigQuery.

Why is DuckDB a good fit for AI agents and LLMs?

Three reasons. It does the arithmetic, so the model never has to guess a total. It runs wherever the agent runs, with no cluster to provision, so an agent can spin up an analytical environment in a second and throw it away. And it now has agent tooling on top: an MCP server, a catalog of SKILL.md playbooks for Claude Code, Codex and Gemini CLI, MotherDuck Guides for business context, and a community extension that lets you write CLAUDE 'what is revenue by product?' directly in SQL.

How accurate is text-to-SQL with DuckDB?

It depends almost entirely on context, not on the engine. dbt's testing found roughly 40% accuracy when a model writes SQL against undecorated tables, 83% with a semantic layer, and near 100% for covered queries with 2026 models. MotherDuck reports similar results with Guides. Treat all vendor figures with care: a CIDR 2026 paper found that 52.8% of the BIRD benchmark and 66.1% of Spider 2.0-Snow labels are themselves wrong. The reliable pattern is a vetted list of queries plus a context layer, with the agent choosing rather than composing.

What are DuckDB's limitations?

One writer per database file (multiple readers are fine); a single machine, so it does not scale out; performance falls off sharply when data is much larger than memory; and no row-level security, GRANT/REVOKE, replication or second-granularity point-in-time recovery. The Quack protocol (beta since 1.5.3, production with 2.0) and DuckLake address the multi-writer problem, and MotherDuck adds the governance and isolation, but plain DuckDB is not a multi-tenant BI platform.

DuckDB vs Databricks, Snowflake or BigQuery: which should a UK business use?

If your analytical data is under a terabyte and your users number in the tens, a DuckDB stack (with MotherDuck or a small API layer, a context layer and an agent) will be faster and an order of magnitude cheaper. If you already run one of the big three, use its native agent (Genie, Cortex Analyst, Conversational Analytics) for governed data and use DuckDB at the edge: in notebooks, pipelines, CI and agent sandboxes. The hybrid is the realistic 2026 answer for mid-market firms.

What is MotherDuck and does it keep data in the UK?

MotherDuck is the managed cloud service built on DuckDB, with a free tier (3 users, 10GB, 10 compute hours a month), a Business plan at $250 a month and per-second compute from $0.60 an hour. It runs in six AWS regions including Frankfurt and Dublin but not London, and EU compute costs around 22% more than US regions. For strict UK residency, self-host DuckDB or DuckLake on a London-region VM or container instead.

What did AWS buying DuckLabs change?

Announced on 26 August 2026 and closing in early September, the deal moves the DuckDB creators and their Amsterdam team into AWS. The project stays MIT-licensed under the independent DuckDB Foundation. AWS already runs 2.5 billion queries through DuckDB inside Amazon Quick, and its stated target is the 90%+ of queries that touch a terabyte or less. Expect an AWS-branded DuckDB service; MotherDuck responded by starting to sell enterprise DuckDB support.

DuckLake is an open lakehouse table format (1.0 in April 2026, MIT) that stores data as Parquet and keeps the metadata in an ordinary SQL database such as Postgres, SQLite or DuckDB, instead of the thousands of small metadata files that Iceberg and Delta write. It gives you snapshots and time travel, which is useful for reproducible agent answers, and multiple writers coordinated through the catalog.

Can DuckDB run in a browser?

Yes. DuckDB-Wasm is the full engine compiled to WebAssembly, about 7MB compressed, with a practical limit of around 2GB of data per tab. The demo in this article runs it in your browser over 5,200 synthetic orders, with no data leaving the page. It suits internal dashboards, embedded analytics and privacy-sensitive exploration.

The Terabyte-or-Less Warehouse: DuckDB, AI Agents and BI Your Team Can Actually Afford

DuckDB is a free, MIT-licensed in-process analytical database (1.5.5, with 2.0 due autumn 2026) whose maker, DuckLabs, AWS agreed to acquire on 26 August 2026, targeting the 90% of queries that run on a terabyte or less. It suits AI agents because the database does the arithmetic, it runs wherever the agent runs, and 2026 tooling (MCP server, 17 SKILL.md skills, MotherDuck Guides, the ACP extension, agent self-signup) gives agents hands and a playbook. Accuracy comes from context, not the engine: about 40% correct on undocumented tables, 83% with a semantic layer, near 100% with a context layer, with a CIDR 2026 caveat that benchmark labels are often wrong. Limits: one writer per file, single node, no row-level security; Quack and DuckLake address multi-writer. The article compares DuckDB and MotherDuck with Databricks Genie, Snowflake Cortex, BigQuery and ClickHouse on scale, concurrency, governance, agent tooling, UK residency and cost, lays out a five-step stack for a UK mid-market firm, and includes a live DuckDB-Wasm demo that runs vetted queries in the browser.

Frequently Asked Questions

What is DuckDB?: DuckDB is a free, MIT-licensed analytical database that runs inside your application process rather than as a separate server. It is columnar and vectorised, so aggregations, joins and window functions over millions of rows run in milliseconds on a laptop, and it reads CSV, Excel, Parquet, JSON, Postgres, S3, Iceberg and DuckLake sources directly without an ETL step. The current stable release is 1.5.5 (July 2026); version 2.0 is due in autumn 2026. Amazon agreed to acquire DuckLabs, the company behind it, on 26 August 2026, with the project staying open source under the independent DuckDB Foundation.
Is DuckDB a data warehouse?: Not on its own. A warehouse is a shared, governed, multi-user service; DuckDB is an engine. It becomes warehouse-like when you add a serving layer (MotherDuck, the Quack client-server protocol, or a small API in front of it), a catalog (DuckLake) and a context layer for agents. For most businesses with a terabyte or less of data and a team you can name, that is enough. For petabytes, hundreds of concurrent users and row-level security, you still want Databricks, Snowflake or BigQuery.
Why is DuckDB a good fit for AI agents and LLMs?: Three reasons. It does the arithmetic, so the model never has to guess a total. It runs wherever the agent runs, with no cluster to provision, so an agent can spin up an analytical environment in a second and throw it away. And it now has agent tooling on top: an MCP server, a catalog of SKILL.md playbooks for Claude Code, Codex and Gemini CLI, MotherDuck Guides for business context, and a community extension that lets you write CLAUDE 'what is revenue by product?' directly in SQL.
How accurate is text-to-SQL with DuckDB?: It depends almost entirely on context, not on the engine. dbt's testing found roughly 40% accuracy when a model writes SQL against undecorated tables, 83% with a semantic layer, and near 100% for covered queries with 2026 models. MotherDuck reports similar results with Guides. Treat all vendor figures with care: a CIDR 2026 paper found that 52.8% of the BIRD benchmark and 66.1% of Spider 2.0-Snow labels are themselves wrong. The reliable pattern is a vetted list of queries plus a context layer, with the agent choosing rather than composing.
What are DuckDB's limitations?: One writer per database file (multiple readers are fine); a single machine, so it does not scale out; performance falls off sharply when data is much larger than memory; and no row-level security, GRANT/REVOKE, replication or second-granularity point-in-time recovery. The Quack protocol (beta since 1.5.3, production with 2.0) and DuckLake address the multi-writer problem, and MotherDuck adds the governance and isolation, but plain DuckDB is not a multi-tenant BI platform.
DuckDB vs Databricks, Snowflake or BigQuery: which should a UK business use?: If your analytical data is under a terabyte and your users number in the tens, a DuckDB stack (with MotherDuck or a small API layer, a context layer and an agent) will be faster and an order of magnitude cheaper. If you already run one of the big three, use its native agent (Genie, Cortex Analyst, Conversational Analytics) for governed data and use DuckDB at the edge: in notebooks, pipelines, CI and agent sandboxes. The hybrid is the realistic 2026 answer for mid-market firms.
What is MotherDuck and does it keep data in the UK?: MotherDuck is the managed cloud service built on DuckDB, with a free tier (3 users, 10GB, 10 compute hours a month), a Business plan at $250 a month and per-second compute from $0.60 an hour. It runs in six AWS regions including Frankfurt and Dublin but not London, and EU compute costs around 22% more than US regions. For strict UK residency, self-host DuckDB or DuckLake on a London-region VM or container instead.
What did AWS buying DuckLabs change?: Announced on 26 August 2026 and closing in early September, the deal moves the DuckDB creators and their Amsterdam team into AWS. The project stays MIT-licensed under the independent DuckDB Foundation. AWS already runs 2.5 billion queries through DuckDB inside Amazon Quick, and its stated target is the 90%+ of queries that touch a terabyte or less. Expect an AWS-branded DuckDB service; MotherDuck responded by starting to sell enterprise DuckDB support.
What is DuckLake?: DuckLake is an open lakehouse table format (1.0 in April 2026, MIT) that stores data as Parquet and keeps the metadata in an ordinary SQL database such as Postgres, SQLite or DuckDB, instead of the thousands of small metadata files that Iceberg and Delta write. It gives you snapshots and time travel, which is useful for reproducible agent answers, and multiple writers coordinated through the catalog.
Can DuckDB run in a browser?: Yes. DuckDB-Wasm is the full engine compiled to WebAssembly, about 7MB compressed, with a practical limit of around 2GB of data per tab. The demo in this article runs it in your browser over 5,200 synthetic orders, with no data leaving the page. It suits internal dashboards, embedded analytics and privacy-sensitive exploration.

Our Services

Contact Cloud First Consulting

Email: info@cloudfirstconsulting.com

Location: London, United Kingdom

Hours: Monday-Friday, 9:00 AM - 6:00 PM GMT

Book a Free 30-min Discovery Call