Changelog

⚖️ Publish and version eval suites​

Did last week’s guide change actually make the agent better? Now you can answer that question without rerunning every test by hand.

Publish an eval suite from the CLI, schedule it to run automatically, and track its performance over time in Context Studio.

Turn the questions you test repeatedly into a versioned suite, then compare results as your guides or semantic models evolve. You’ll quickly see what improved agent performance or get an earlier warning when a change causes a regression.

Publishing and scheduling eval suites requires a Manager or Admin workspace role. Learn how scheduled evals work.

✅ Skip agent approvals​

No one wants to babysit their agent or come back to find it waiting for approval. Now you can let the Hex Agent run in YOLO mode (responsibly, of course).

There are two options in the prompt bar when building with the Hex Agent, you can either:

• Stick with Ask for approval mode and the agent will pause for confirmation before destructive actions like deleting a cell or publishing.
• Choose Skip approvals mode and the agent works straight through without pausing for confirmation on every edit. Don’t worry, you can still undo changes from the last agent turn.

Anything that expands access like granting OAuth, reconnecting an app, or authorizing an MCP server still stops and waits for your approval.

🛠️ Your chat agent can build in Hex​

Your chat agent can do a lot more than ask Hex questions.

Once you connect Hex to an MCP client like Claude, ChatGPT, or Cursor, your agent can create, retrieve, and update Hex projects with the create_project, get_project, and update_project tools.

That means you can stay in your chat agent to:

• Investigate what happened and why with an ad hoc analysis
• Plan what happens next with a forecast or scenario model
• Turn a recurring workflow into an interactive data app

Project editing requires an Editor role or higher and the appropriate project permissions. Changes made through MCP apply directly to a project’s draft notebook, so they do not update its published app version.

Learn more about editing Hex projects via MCP.

🎮 More control over Generative app iteration​

Directly edit text, create new chart properties, and queue up viz tweaks all in line.

✏️ Direct text editing​

You can also edit text directly in the app builder, just like you would in a doc. It’s perfect for tweaking titles, or updating copy in your data apps and you can now view the underlying source code with just 1 click.

If your text has a computed variables, you can inline prompt the agent to make updates — this is particularly helpful for things like quick formatting changes.

📊 Targeted Chart Updates​

Reorganize charts or adjust a visualization without touching the rest of the app. Select the chart and ask the agent to switch the chart type, update the data, or refine it’s layout — without regenerating everything around it.

💬 Create new chart properties with a prompt​

Need a control that’s not already on the Style panel? You can now select a component, describe the property you want, and Hex creates it.

Once added, you can adjust that property directly alongside the charts other controls.

📎 Analyze CSVs directly in Slack​

You can now upload a CSV in Slack and ask Hex to analyze it without leaving the conversation. Attach a file, mention @Hex, and ask for a summary, trends, top performers, anomalies, or anything else you want to explore.

Hex returns key takeaways and tables directly in the thread, plus a link to continue working in Hex.

🔑 Clearer OAuth errors in published apps​

When a publisher’s OAuth credentials expire, viewers can still access the published app in a read-only state instead of being blocked by an error screen. The publisher receives a clear error message to refresh their token and restore live data.

Expired tokens are also easier to identify in connected app settings, and SQL cells now explain when credentials are the reason a query can’t run.

🔐 Bring your own AI provider​

AWS Bedrock and Google Gemini Enterprise Agent Platform are now available as bring-your-own-key providers in Hex.

If your organization already manages AI through AWS or Google Cloud, you can bring that setup to Hex. Claude inference runs through your provider account, so your existing security controls, agreements, and usage logs stay in place.

Already route AI traffic through a compatible third-party gateway? You can keep using that infrastructure with Hex, too.

Admins can get started in Settings → AI & agents → Provider settings → Add an API key.

Learn more about BYOK.

🔎 More visibility and model choice for BYOK​

It’s now easier to track who’s using AI (and choose the right model) when you bring your own key.

Every BYOK request includes the Hex user’s email, so admins can attribute usage, split costs across teams, and investigate activity in their AI gateway.

Organizations using an OpenAI key can also choose Sol, Terra, or Luna. Pick the model that fits the job without switching away from BYOK.

🎨 Generative chart controls​

Previously, even small edits to your charts in generative apps required new prompts to the agent. Now, you can build custom, code-generated apps with the Hex Agent, and then use the Style panel to make point-and-click chart polishes.

Some edits you can now make this way:

• Update the chart title, style and height
• Restyle series and change the colors
• Rename axes
• Control legend titles, visibility, and placement

This makes small edits faster, and saves your valuable tokens for bigger changes that require deeper thinking.

Try this out on an app you’ve already built or start building a new one with just a single prompt.

🔐 Shareable results for OAuth connections​

Orgs with OAuth connections can collaborate, and share results, without needing to share credentials.

Importantly, this unlocks threads sharing, better performance through shared query caching, and more flexible collaboration.

To take advantage of this behavior, Admins should set Authoring queries to Require personal credentials and set Sharing results to Shared across users.

Existing connections keep results private to each user to prevent unintended data exposure. When results are private, agent threads remain isolated to the user who generated them. This setting applies per data connection, not per project.

⌨️ Semantic projects via CLI​

Pull your semantic spec out of Hex and into your repo, for local development and source control. This makes it easier to audit context or track what's changed.

Read and export semantic projects from the command line with the 'hex context semantic-project' command group

• 'list' — see all semantic projects in the workspace
• 'get' — get a project, with '--include-fields' for dimensions, measures, and joins
• 'model get' and 'view get' — pull a single model or view
• 'pull' — write the source files of a project's latest version to a local directory

To bring semantic projects into your development workflow, start by installing the Hex CLI.

🤖 Quickly toggle from chat to notebook​

Previously, you needed to wait a minute for Hex to convert your Thread to a notebook, now it happens automatically.

After starting a conversation with the Hex Agent, you can easily switch views and build in the underlying notebook. And if you're in the notebook view but just want to focus on the conversation, swap views so the code stays out of the way.

The conversation persists no matter what view you're toggling between.

Worried about your conversations cluttering your project list? Not to worry — projects created from agent conversations are unlisted to start, so nothing shows up in your project list until you decide to promote it or save it as a published app.

💬 The Hex Agent reads the whole Slack thread​

Tag @Hex mid-conversation and the Hex Agent now picks up the discussion that came before it.

Previously, the agent only saw the message that mentioned it, so "can you dig into this?" or "can you check these numbers" meant you had to restate the question. Now the Hex Agent pulls in the original message and replies, along with who said what.

That means threaded follow-ups actually get answered instead of met with "which metric do you mean?", and handoffs from a team discussion to the agent don't require restating what was already said.

Learn more about the Hex Agent in Slack.