Steward approvals and tool permissions¶
By default, Steward asks before it changes your semantic layer or an AI Analyst, or runs a SQL query against your warehouse. It shows an approval card in the chat, and nothing happens until you accept. Saving to memory doesn't need approval.
You can change this per Workspace in Tool permissions: let some tools run without asking, or block them.
Availability
- Who can do this: members with the Edit capability in the active Workspace.
Review before you approve
Steward can make mistakes. Review its proposed edits before you approve them. See Accuracy and limitations.
What needs your approval¶
With the built-in settings, Steward asks before it uses these tools:
| Tool (as shown in the app) | What it does |
|---|---|
| Apply edits | Applies changes to your semantic layer. |
| Write AI Analyst | Creates an AI Analyst or changes one. |
| Run SQL query | Runs a read-only SQL query against one of your data connections. |
Everything else Steward does, such as searching the semantic layer, listing models and metrics, inspecting warehouse tables or reading AI Analyst conversations, runs without an approval card. Plan mode and questions from Steward use their own cards.
Review an approval card¶
An approval card appears in the Steward chat when Steward is waiting for your decision. It shows:
- A header such as Steward wants to apply edits, Steward wants to write AI Analyst or Steward wants to run SQL query.
- Steward's explanation of what it wants to do and why.
- A one-line summary, for example
Apply 3 edits: …. - View details, which expands the full list of changes. Click Hide details to collapse it again.
Semantic-layer edits¶
Under View details, each proposed edit is listed by what it does, for example Create model or Update measure. For each field that changes, you see the Before and After values. An edit that changes only metadata shows Metadata only — no field changes to preview.
Steward can create, update and delete models, dimensions, measures, filters, relationships, metrics and glossary terms. It can also restore deleted models, metrics and glossary terms, and add or remove glossary term mappings.
Before the card appears, Steward test-runs the proposed expressions against your warehouse:
- Edits that fail this check are left off the card, and Steward goes back to fix them.
- If Steward can't reach the data connection, or every edit fails the check, no card appears and nothing is applied.
- If Steward can't find the target of an edit (for example, it was deleted), the card says how many edits couldn't be previewed and lists them under Couldn't be previewed. These edits will likely be rejected. The rest still apply.
One card holds at most 50 edits. Each edit is applied on its own, so one failed edit doesn't stop the others.
AI Analyst changes¶
When Steward creates or changes an AI Analyst, one card covers every part of the change. Depending on the change, View details shows:
| Section | What it shows |
|---|---|
| New analyst | Steward is creating a new AI Analyst. |
| Identity | Changes to Name and Description. |
| Proposed instructions diff | The lines Steward wants to add to or remove from the analyst's instructions. |
| Models to connect / Models to disconnect | Models Steward wants to give the analyst access to, or take away. |
| Glossary access | Changes to the glossary terms the analyst can use. |
| Suggestions | New suggestion cards for the analyst's Explorer homepage. They replace the existing cards in the categories shown. |
If the analyst gets its models through tags, connecting a model directly switches how model access works. The card warns you: models the analyst reaches through tags today stay connected, but newly tagged models are no longer added automatically.
Steward never deploys or undeploys an AI Analyst, and never changes who can use it in Explorer.
SQL queries¶
Steward can ask to run a SQL query against one of your data connections, for example to check what a column contains. The card shows Steward's reason for the query and the line Read-only SQL query on "<data source>" — it cannot change any data. The SQL itself is under View details.
Steward accepts only a single SELECT statement (or WITH … SELECT). Statements that write data or change the schema, and multiple statements, are rejected before anything runs. Results are capped at 200 rows by default, and at most 1,000 rows and 50 columns. Approving a query lets Steward read the returned rows, so check the statement and the data connection before you accept.
Accept or request changes¶
Each card has these buttons:
| Button | What happens |
|---|---|
| Accept | Steward goes ahead with this one action. |
| Accept, don't ask again in this chat | Steward goes ahead, and won't ask again for this tool for the rest of this chat. Other chats and the Workspace setting don't change. |
| Request changes | Nothing is applied or run. Steward revises its proposal or asks you what to change. |
If Accept, don't ask again in this chat can't save your preference, you see Couldn't save that preference. Try again, or accept just this once. Click Accept to approve only this card.
Tip
For larger changes, turn on Plan mode in the composer. Steward then drafts a plan for you to review before it changes anything. See Using Steward.
Set tool permissions for a Workspace¶
Tool permissions apply to everyone using Steward in the active Workspace. Each Workspace has its own settings.
- In Studio, open the Steward panel.
- Click Steward settings (the gear icon), then select Tools.
- In Tool permissions, choose a mode for each tool:
- Always allow: Steward uses the tool without asking.
- Ask first: Steward shows an approval card first.
- Always deny: Steward can't use the tool.
Each change saves straight away and shows Saved.
Tools are grouped into Data access (read-only tools that explore your semantic layer and Workspace), Sensitive (tools that change the semantic layer, instructions or memory) and Other. Use Allow all or Ask all to set every tool in a group at once.
Always allow removes the review step
With Always allow on Apply edits, Write AI Analyst or Run SQL query, Steward makes those changes or runs those queries without showing a card, for everyone using Steward in this Workspace. Each change saves straight away, without a confirmation. Allow all in the Sensitive group applies this to every tool in the group. You can still undo semantic-layer and AI Analyst changes. See Rolling back Steward changes.
Under each tool, the dialog shows where its current mode comes from:
- Built-in default: nobody has changed this tool in this Workspace.
- Environment default (here, "environment" means the Workspace): someone set this mode for the Workspace in this dialog.
The Tools and Memory items appear in the Steward settings menu only for members with the Edit capability.
Which settings take effect¶
| Tool | Built-in default | What the modes do |
|---|---|---|
| Apply edits | Ask first | All three modes work as described. |
| Write AI Analyst | Ask first | All three modes work as described. |
| Run SQL query | Ask first | All three modes work as described. |
| Save to memory | Always allow | Steward never shows a card for memory. Ask first behaves like Always allow. Only Always deny changes anything: it stops Steward from saving to memory. |
| Other tools in the dialog | Always allow | These are read-only tools, such as Search semantic layer, List models and Inspect warehouse tables. In the current release, Steward runs them whatever mode you choose, so don't rely on Always deny to stop Steward searching the web or reading attachments. |
When a tool is set to Always deny, Steward doesn't show a card. It is told the tool is blocked, nothing is applied, and it looks for another way or explains what it couldn't do.
For more about Steward's memory, see Memory and saved prompts.
How settings combine¶
When Steward is about to use a tool, it checks, in this order:
- A per-chat choice made with Accept, don't ask again in this chat.
- The Workspace setting from Tool permissions.
- The built-in default.
The first one it finds wins. Per-chat choices aren't shown in the Tool permissions dialog and don't affect other chats.
Audit log¶
Every change to a tool permission is recorded in the audit log as steward.tool_permission_set. Choosing the mode a tool already has records nothing. See Audit logs.
Troubleshooting¶
I don't see Tools in the Steward settings menu. You need the Edit capability in the active Workspace.
Steward said it couldn't make a change, and no card appeared. The tool may be set to Always deny in this Workspace. Check Steward settings → Tools. If the tool is allowed, Steward may not have been able to reach the data connection to check its edits. Make sure the connection works, then ask Steward to try again.
I accepted a change by mistake. You can undo Steward's changes from the chat. See Rolling back Steward changes.