Hex

A hosted SQL and Python notebook whose cells form a reactive graph and can be published as data apps. It now leads with AI agents that answer questions from semantic models and curated context.

What it is: Data notebooks,Semantic layers
Type: Commercial product
Made by: Hex Technologies Inc.
License: Proprietary
Facts checked: 9 October 2026

What it is

A Hex project is a notebook of SQL, Python, Markdown, chart, pivot, input and other cells that runs in Hex’s cloud. A SQL cell queries a warehouse connection in that warehouse’s own dialect, or queries dataframes and uploaded CSVs with DuckDB (“dataframe SQL”). Its result becomes a variable: by default a pandas dataframe that Python cells can use, or, in query mode, a query of which Hex fetches only 1,000 preview rows and inlines the full SQL as a CTE into downstream warehouse queries (“chained SQL”), so big results stay in the warehouse. Python cells run in a kernel that Hex starts for each notebook, app, scheduled or API session.

Unlike a classic notebook, the project is a graph. Hex infers the links between cells from variable references in Python and SQL, and in its default Auto mode a run updates every downstream and stale cell; running a single cell leaves its dependants marked stale. The same graph powers apps: you pick the inputs, text and output cells to show, publish a version separate from the draft you keep editing, and an input re-runs only the logic downstream of it, while cells the app doesn’t need are skipped. Scheduled runs (at most hourly) and the REST API rerun a published app and can refresh what viewers see.

Hex now positions itself as AI analytics. The Notebook Agent (August 2025) plans and writes chains of SQL and Python cells; Threads (October 2025) lets people who don’t code ask questions in chat, drawing on semantic models, endorsed tables, workspace rules and published apps. Semantic models are authored as YAML in Hex’s Modeling Workbench (a public beta) or synced from dbt MetricFlow, Cube or Snowflake semantic views. Since July 2026, AI use is metered in credits. Hex is independent: it acquired the BI startup Hashboard in April 2025, raised a $70 million Series C in May 2025, and says it has over 2,000 customers.

Why teams pick it

SQL and Python in one project: warehouse SQL, DuckDB dataframe SQL and pandas pass results to each other as variables, and chained SQL keeps large results in the warehouse
Reactive execution inferred from the code, so re-runs are predictable and an app’s inputs recompute only the cells downstream of them
One project goes from exploration to a published app, with scheduled runs, notifications and a REST API that Airflow, Dagster or your own code can call
Agents grounded in the workspace: the Notebook Agent and Threads use semantic models (built in Hex or synced from dbt, Cube or Snowflake), endorsed tables and earlier published work

Watch out for

Closed SaaS: projects and kernels run in Hex’s cloud (multi-tenant by default, with EU, single-tenant and HIPAA options), and projects live in Hex rather than in your repo, though they can be exported to Git
Reactivity relies on spotting variable references, so variables created through globals() can break re-runs; kernels have hard limits: 25 concurrent per user, a 15-minute idle timeout for app sessions, 60 minutes for scheduled and API runs, 36 hours at most
Priced per editor ($36 a month on Professional, $75 on Team, Enterprise on request), and since July 2026 AI features consume credits; the monthly per-seat grants don’t roll over
Much is plan-gated or in beta: scheduled runs and cron schedules need Team (Professional gets one scheduled run per project), semantic authoring is a public beta on Team and Enterprise, and Databricks metric views are in beta without sync

How it fits with other technologies

What Hex uses

Run a published Hex project with new inputs through the REST API, then check its status · bash

# Trigger a run of a published project with new input values
curl -s -X POST "https://app.hex.tech/api/v1/projects/$PROJECT_ID/runs" \
  -H "Authorization: Bearer $HEX_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"inputParams": {"region": "EMEA", "min_orders": 10}, "useCachedSqlResults": false}' \
  | tee run.json | jq -r .runUrl              # a snapshot of this run's results

# Poll the run: PENDING, RUNNING, then COMPLETED or ERRORED
curl -s "https://app.hex.tech/api/v1/projects/$PROJECT_ID/runs/$(jq -r .runId run.json)" \
  -H "Authorization: Bearer $HEX_TOKEN" | jq -r .status

Plans#

From Hex’s pricing page, checked 2026-10-09. Prices are per editor per month.

	Community	Professional	Team	Enterprise
Price	Free	$36	$75 (14-day free trial)	Custom
Published apps	Up to 5	Up to 5	Unlimited	Unlimited
AI	Hex agent trial	Hex agent, standard credits	Extended credits	Premium credits
Version history	7 days	30 days	Unlimited	Unlimited
Compute	Small (4 GB, 0.5 CPU)	Medium (8 GB, 1 CPU)	Medium, plus pay-as-you-go profiles	Medium, plus pay-as-you-go profiles
Also	All cell types, any data source		Scheduled runs and alerts, semantic models, shared components	Audit logs, OIDC SSO, BYOK, single-tenant and HIPAA add-ons

Larger compute is billed per minute: from $0.32 an hour for Large (16 GB, 2 CPUs) to $2.58 for 4XL (128 GB, 16 CPUs), and $2.93 to $4.06 an hour for an L4 or A10G GPU. Students and educators get Professional free.

See it in a real stack

No architecture uses Hex yet. Drop it into the playground and wire it up yourself.Open playground →

Where it came from2020–2026

Barry McCardel, Caitlin Colgrove and Glen Takahashi met while working at Palantir. They started Hex around the turn of 2020 to build the data tool they had always wanted: as analysts and data scientists they jumped between local Python notebooks, BI tools, spreadsheets, SQL scratchpads and scripts, and shared results as copy-pasted charts, emailed CSVs and one-off decks. A $5.5 million seed round led by Amplify Partners followed in 2020, and in March 2021 Hex opened up as a “Data Workspace for teams”: SQL and Python notebooks with real-time multiplayer editing, and an App Builder to turn an analysis into something stakeholders could click through instead of asking for one more cut of the data.

In October 2021 Hex 2.0 replaced top-to-bottom execution with a reactive graph, and Hex announced a $16 million Series A led by Tomasz Tunguz of Redpoint, along with a first dbt integration. A $52 million Series B led by Andreessen Horowitz, with Snowflake and Databricks investing, followed in March 2022, and a $28 million round led by Sequoia in March 2023 arrived with Magic, AI features that wrote SQL from plain language and by 2024 could generate several chained cells at once.

AI then became the company’s centre. In April 2025 Hex acquired Hashboard, whose self-serve BI, semantic modeling and visualization work it set out to bring into Hex, and in May 2025 it raised a $70 million Series C led by Avra, saying over 1,500 teams used it. The Notebook Agent arrived in August 2025 and Threads, conversational analytics introduced by Hashboard’s founder Carlos Aguilar, in October 2025. By July 2026 Hex said agents had overtaken manual cell creation as the main way people worked in it, and moved AI use onto credits.

Why it’s built this way

SQL cells as first-class citizens beside Python: Hex’s founders saw most data people as “analytically technical”: technical in how they work, but not software engineers who want to manage environments. Built-in SQL cells with templating and caching, and later dataframe SQL, answered “Python or SQL?” with both.
A reactive graph instead of a linear notebook: Shared mutable state in imperative notebooks hurts interpretability, reproducibility and performance. Hex parses each cell’s code, builds a DAG and re-runs dependants, which also lets a published app recompute only what an input touches.
Apps published from the same project: Sharing analysis meant screenshots in decks and long message threads. Publishing selected cells as an app, with inputs such as a checkbox or dropdown, lets stakeholders explore without a separate dashboard tool.
Agents built into the workspace, not a standalone chatbot: Hashboard had tried wiring an LLM to semantic models and found it didn’t work well enough. Threads reuses the Notebook Agent’s framework and the workspace’s context, and hands off to a notebook when the data team needs to inspect or extend an answer.

Timeline

2020
$5.5M seed round led by Amplify Partners
2021
Public launch in March; Hex 2.0 with a reactive graph and a $16M Series A in October
2022
$52M Series B led by Andreessen Horowitz, with Snowflake and Databricks
2023
$28M round led by Sequoia; Magic AI features
2025
Acquires Hashboard in April; $70M Series C in May
2025
Notebook Agent in August, Threads in October
2026
AI usage moves to credits in July

Sources17 · facts checked 9 October 2026

Improve this page on GitHub