Hex

The Hex agent writes and runs SQL and Python against Snowflake, BigQuery, Databricks, Redshift, Postgres and other connected warehouses, and as an MCP client it calls tools on Notion, Linear or any other MCP server the workspace connects (public beta).

The agent works with any data connection or semantic model the user already has access to, and admins can let it search the web and extract content from URLs. Team plans add the REST APIs, OAuth and integrations with Airflow, Dagster and dbt, and Enterprise adds OAuth database connections. Users can also upload CSV files or images for the agent to work with, and an uploaded CSV can be joined with warehouse data.

SourceHex, learn.hex.tech Notebook agent docs and changelog, and hex.tech/pricingread 2026-10-05

Work is built as notebooks of SQL, Python and no code cells that the team arranges and the agent creates, edits and runs cell by cell, and a whole notebook runs as one flow across its cells. The Notebook Agent can add Python, SQL, Markdown, pivot, input parameter, single value and chart cells, move cells around and delete them with the dependent cells in view, and lay out the published app in rows, columns and tabs.

Fix with agent takes on multistep repairs across cells, while Quick fix handles small errors such as syntax. A chat agent in Claude, ChatGPT or Cursor can also create a Hex project through the MCP server, add cells and run the whole notebook, with its edits landing in the draft rather than the published app.

SourceHex, learn.hex.tech Notebook agent and MCP server docsread 2026-10-05

Agents ground answers in the organization's context: they search endorsed semantic models synced from dbt, and Context Studio manages the semantic models, guides and skills they draw on. Guides take two forms, a workspace context file that every conversation reads and a library the agent pulls from by relevance, and a guide can point the agent at specific projects.

Admins and managers edit them together in Context Studio, test questions against a draft before publishing and can roll back to an earlier version, or manage them from the CLI or GitHub Actions, in which case they are read only inside Hex. Semantic model views let a data team rename and regroup dimensions and measures for business users, and the Notebook Agent also reads other projects the user can explore.

SourceHex, learn.hex.tech Guides and Notebook agent docs and changelogread 2026-10-05

By default the agent runs in Ask for approval mode and pauses before destructive actions such as deleting cells or publishing; a Skip approvals mode exists, the last agent turn can be undone, and security sensitive actions such as OAuth authorization always require approval. In approval mode, each cell the agent changes waits for the user to confirm or undo it.

Admins can mark data connections as sensitive, and a client reaching a sensitive connection through the MCP server needs admin approval first. For the Slack agent, admins choose whether answers drawn from sensitive data arrive as a private reply, as a link only, or not at all.

SourceHex, learn.hex.tech changelog, Notebook agent, MCP server and Hex Agent in Slack docsread 2026-10-05

Hex holds a SOC 2 Type II report and lists HIPAA, PCI and GDPR in its trust center, supports SSO through Google, Okta and OIDC, limits actions by user role, and offers audit logs and BYOK on Enterprise. Data at rest, including database credentials, uploaded files and cached query results, is encrypted with AES 256, and traffic uses TLS 1.2 or newer.

Permissions are set across user roles, data access and project access. Production systems at Hex sit behind multifactor authentication, and the company runs regular penetration tests, a private bug bounty program and code scanning. OIDC SSO is listed on the Enterprise plan, and HIPAA coverage is sold there as an add on.

SourceHex, hex.tech/security and /pricingread 2026-10-05

The agent's actions are the notebook cells it writes and runs, visible in the project with their code and outputs, Context Studio shows which model the agent picked and thread level token spend, and Enterprise adds an Observability API. In Context Studio, admins and managers on Team and Enterprise track agent conversation volume and the most active users, and admins can open any conversation to see which context the agent drew on, while managers see totals only. Scheduled run history is kept for 90 days.

SourceHex, learn.hex.tech changelog and Scheduled runs docs, and hex.tech/pricingread 2026-10-05

Threads persist and can be continued later, and notebooks keep version history. Hex describes no memory that the agent keeps from one conversation to the next, and guides in Context Studio are context that people write for the agent. Notebook Agent conversation history lasts 30 days, and starting a new thread clears the context. Each user can also set personal instructions for the agent in a rules file.

SourceHex, learn.hex.tech Notebook agent docs and changelogread 2026-10-05

Hex can be deployed single tenant in a geographic region of the customer's choosing among those AWS supports, sold as an Enterprise add on, alongside the multi tenant cloud on AWS. Custom deployments run on their own address, such as eu.hex.tech or a company subdomain, which takes the place of app.hex.tech for tools like the MCP server. Team and Enterprise workspaces can add larger compute profiles billed by the minute, from 16 GB of memory up to 128 GB and A10G or L4 GPUs, while the free Community plan runs on small compute and Professional on medium.

SourceHex, hex.tech/security and /pricing, and learn.hex.tech MCP server docsread 2026-10-05

A public gallery of more than 40 ready made analysis templates across 13 categories, such as customer segmentation, anomaly detection, A/B testing and a text to SQL chatbot, ships alongside the built in Notebook Agent and Threads.

The categories range from clustering and feature selection to KPI dashboards, natural language processing, time series and Snowpark, with examples such as market basket analysis, collaborative filtering and geospatial analysis.

Teams can also package reusable cells as shared components and group work into shared collections, with three of each on the free plan and unlimited collections on Team.

SourceHex, hex.tech/templates and /pricingread 2026-10-05

Scheduled runs and alerts on Team and Enterprise run projects without a person starting them, and the agent is also reachable in Slack, through MCP clients, the CLI and the APIs.

On those plans schedules run hourly, daily, weekly, monthly or on a cron expression, and conditional notifications go out only when a set condition is met, while Professional allows one scheduled run per project without Slack notifications or custom schedules.

Tasks, in beta, put the Threads agent on a daily, weekly or monthly schedule and deliver its answer to Slack or email with a link back to the Thread. In Slack, people tag @Hex in a channel or thread or message the app directly.

SourceHex, learn.hex.tech Scheduled runs, Tasks and Hex Agent in Slack docs, and hex.tech/pricingread 2026-10-05

Users pick a model in the agent's model picker or leave it on Auto, which selects the model suited to the task, across Anthropic and OpenAI models served directly or through AWS Bedrock and Google; admins can bring their own provider keys so inference runs through their own accounts. Effort levels can be set beside the model choice.

Bring your own key is an Enterprise feature that covers Anthropic, OpenAI, AWS Bedrock and Google Gemini Enterprise Agent Platform, and admins can make it mandatory. By default the providers run under zero data retention and do not train on customer data, and admins decide whether to allow models that need temporary retention for safety monitoring. Agent use draws on monthly credits per seat, standard on Professional, extended on Team and premium on Enterprise, and the free plan includes an agent trial.

SourceHex, learn.hex.tech AI data privacy and Notebook agent docs, and hex.tech/pricingread 2026-10-05

An MCP server at app.hex.tech/mcp with OAuth exposes named tools to search projects, create and continue Threads and create, edit, delete and run notebook cells, alongside a public REST API and a CLI. The server runs on Team and Enterprise, where tools that search projects and run Threads need the Explorer role and tools that change or run cells need Editor or above. Claude Desktop, Claude Code, Cursor, ChatGPT, Codex, Glean and Figma are named as clients. The REST APIs come with Team, and Enterprise adds an Observability API.

SourceHex, learn.hex.tech MCP server docs and hex.tech/pricingread 2026-10-05

Teams turn the questions they test repeatedly into versioned eval suites, publish them from the CLI, schedule them to run and track agent performance over time in Context Studio, catching regressions when guides or semantic models change. Publishing and scheduling a suite takes the Manager or Admin role. Guide changes can be previewed and tested against specific questions before they go live, and each published version stays in a history that can be compared and rolled back.

SourceHex, learn.hex.tech changelog and Guides docsread 2026-10-05

Hex does not describe the agent operating a browser or desktop; it works through code, notebook tools and MCP. When an admin turns it on, the agent can also search the web and pull content from a URL.

SourceHex, learn.hex.tech Notebook agent docs; hex.techread 2026-10-05