AI configuration

Use AI Configuration to connect DataMind OS to the Kartlos AI service, choose which AI models people in your company may use, and manage the skills your agents load. The page has two tabs: Connection & models and Skills. Everything on the Skills tab is described in Organisation skills.

Set the connection up once, then come back whenever you want to change the models on offer or the skills that reach your agents.

Before you start

• The Connection & models tab needs AI Configuration Viewer to see the settings, and AI Configuration Manager to change them. Without the manager right, every field and switch is read-only.
• The Skills tab needs AI Skills Viewer to see the skills, and AI Skills Manager to change them. Without the manager right, the skills are read-only.

The Admin Viewer and Admin Configurator High-Level presets include these. You see only the tabs your permissions allow: with only the skills permissions, the page shows the skills alone, without the Connection & models tab. See the Permission reference.

The installation's key is its License Key, entered in Admin > License (License). AI Configuration shows it read-only as Platform key, with Managed in Admin > License beside it.

Connect the AI service

1. Open Admin in the sidebar, choose AI Configuration at the top of the page, and open the Connection & models tab.
2. In the AI connection panel, check Platform key: it shows the start and end of the installation's key, such as dm-lic-8ai5nefy••••3kq9. If it reads No key yet: enter a License Key in Admin > License, enter the key on License first.
3. Keep the address that API host shows. Change it only if DataMind gives you a different one, and save the change before you test it.
4. Click Test Connection. It checks the saved key against the saved API host.
5. Check that the switch next to the connection reads Enabled. A new connection starts switched on.
6. In Allowed models, switch on each model your company may use.
7. Click Save in the bar at the bottom of the page.

API host is required. If you save without it, the field shows API Host is required.

After the first save, a confirmation reads "AI model settings added successfully". Later saves read "AI model settings updated successfully".

Settings and defaults

Choose the models people can use

Models are opt-in. Only the models you switch on here can be used anywhere in DataMind OS, including chat pickers and Mind settings. Changes take effect as soon as you save.

Use the Allowed models panel like this:

• The count next to Allowed models shows how many models are on, out of the total. It never drops below two, because the built-in Kartlos models are always on.
• Type in Search models… to find a model by its name or ID.
• Use the filters All, Allowed and Off to show every model, only the models that are on, or only the models that are off.
• Hover over a model's name to see its context length and what it can do: Standard or Premium, Reasoning, Tools, Structured and Image.
• Read the badges. New marks a recently released model. Frontier, Pro or Lite shows the model's category, where one is assigned. Always on marks the two built-in Kartlos models, whose switches cannot be changed. One model also carries a Premium badge.

A newly released model arrives switched off. Switch it on to allow people to use it.

Privacy and policy

The Privacy & policy section sits under the connection panel and has two cards.

• Zero Completion Insurance, under Billing protection: you are not charged for a response that has no output tokens and ends with a blank finish reason or an error. It is enabled automatically for all accounts.
• Zero Data Retention, under Data privacy: your data is never stored on external servers, and it is deleted immediately after serving its purpose.

To protect personal data that people type into AI, see PII detection.

Test the connection

The result of Test Connection appears under the button.

• Connection test successful: <API host> appears in green when the key and address work.
• A message in red appears when they do not, starting with the address it tested. It comes from the AI service and says why the test failed. If the service gives no reason, you see Connection failed.

The test always uses the saved key and the saved API host. While the field shows another address than the saved one, the button is off and reads Save the API host to test it; save the new address first. While a test is running, or while a save is in progress, the button is disabled. A failed test is not retried or cached: fix what the message points to, then test again.

Limits and edge cases

• Adding a model needs the live model catalogue. If the catalogue is unavailable, the Allowed models panel shows Model catalog unavailable with The live catalog could not be reached. You can still switch models off, but you cannot switch new ones on until it recovers. Click Retry to check again.
• Saving fails if you add a model that the catalogue does not list. The error names the model. If the catalogue cannot be reached at all, a message starting with Model catalog is unavailable explains that the list cannot be validated and asks you to try again shortly.
• A chat or Mind that uses a model you have switched off is refused. The person sees a message asking them to pick a different model. Mind settings follow the allowed list, so a Mind can only use the models you have switched on. To choose a model for one Mind, see Choose models and AI behaviour.
• Without edit permission, every field and switch is read-only, the Test Connection button is not shown, and no save bar appears.
• When there are unsaved changes, the bar reads You have unsaved changes. Click Cancel to go back to the last saved settings.
• If a save fails on an existing setup, the error appears under the connection panel.
• If the panel has nothing to show, it reads No models match your search. or No models in the catalog.
• Other messages you may see are listed in Administration troubleshooting.

Manage the skills agents use

The Skills tab lists every skill your agents can load, from three sources. The filter above the list switches between them:

• Organization: skills your administrators wrote and push to the agents you pick and to your Oda users.
• Built-in: skills that ship with DataMind.
• Mind skills: skills that a Mind owns.

On Organization, each row shows the skill's Name, the Agents it reaches, whether it is Forced, its Roles, its Revision, its Last test, its Usage and the date it was Updated. Last test shows the result of the skill's newest benchmark, and Usage how often assistants loaded it in the last 30 days. Mind skills also shows Last test and Usage. Use Search by name or title to narrow the list. When the list is longer than the page shows, it says how many of the total are shown and asks you to search to narrow it.

Create an organization skill

Open the editor for a new organization skill and fill its fields:

• Name: the command people run. The name is fixed: Oda writes it as a folder and users run it as a slash command.
• Title: the name the list shows for the skill.
• Description: agents decide from this text whether to load the skill, so name the tasks it is for.
• When to use: guidance on the requests that should trigger it.
• Agents: pick every agent the skill reaches. The choices are Oda, Analytical Minds, Warehouse Builder Minds, Lake Builder Minds, Model AI, Lake connector assistants, Slides AI and Custom Visual AI.
• Delivery to Oda: Managed ("Your edits reach everyone. Removing it removes it from everyone.") or Install once ("Each user gets a copy they own. Later edits don't reach them.").
• Forced: turn it on so "Every agent you picked lists it and nobody can turn it off. Needs managed delivery and cannot be command only.".
• Slash command and Only with /{name}: the command people type, and whether an agent runs the skill only when someone types it.
• Only for specific roles: restrict the skill to the roles you pick.

For the steps, the limits and the names you cannot use, see Create an organisation skill.

An empty list reads "No organization skills yet" over "Organization skills reach the agents you pick, and Oda users. Create one or import a SKILL.md file.".

Import a skill

Import reads a SKILL.md file or a .zip into a draft. Nothing is saved until you save the skill.

History and promotion

Every save keeps a revision. History lists the revisions of a skill, shows the differences between two of them, and lets you restore an older revision as a new one.

Promote to organization turns a Mind skill into an organization skill: "Creates an organization skill from this Mind skill. The Mind keeps its own copy.". You cannot promote a skill you wrote or edited yourself. The message reads "You wrote or edited this skill, so another administrator has to promote it.", and another administrator has to promote it.

The revision you promote also needs a passing benchmark, run in the Mind that owns the skill. See Promote a Mind skill to the organisation and Test a skill.

Remove an organization skill

Deleting a skill asks Delete "{name}"? and warns that agents and Oda users stop getting it within minutes and that the name stays reserved.

Questions, answered

Where do I change the key?

In Admin > License. AI Configuration shows the key read-only as Platform key, with Managed in Admin > License beside it. See License.

What does a failed Test Connection message mean?

It is the reason the AI service gave for refusing the connection, after the address that was tested. It appears in red under Test Connection, and it can be a technical sentence from the service. Check the key on License and the saved API host, then test again.

A new model was released, but nobody can pick it. Why?

New models arrive switched off. Switch the model on in Allowed models, then click Save.

Can I switch off the built-in Kartlos models?

No. The built-in Kartlos models are always on, and their switches cannot be changed.

I switched off a model. What happens to chats and Minds that use it?

Requests that use the model are refused, and the person sees a message asking them to pick a different model. Mind settings follow the allowed list, so a Mind can only use the models you have switched on. To choose a model for one Mind, see Choose models and AI behaviour.

What do I do when the model list says "Model catalog unavailable"?

Click Retry. Until the catalogue is reachable, you can switch models off but not on.

Am I charged for empty AI answers?

Not in the case the Zero Completion Insurance card describes: a response with no output tokens that ends with a blank finish reason or an error.

Who can change the AI settings?

Anyone with the AI Configuration Manager permission. People with only the AI Configuration Viewer permission can see the settings, but they cannot change them.

Where are the skills I pushed to Oda users now?

On the Skills tab of AI Configuration. The Oda page keeps General, MCP and Feedback, and it says that Oda skills are managed in AI Configuration.

Who can see the Skills tab?

Anyone with the AI Skills Viewer permission, and anyone with AI Skills Manager, who can also change the skills. Admins without either right see only the Connection & models tab. With a skills permission but without AI Configuration Viewer, the page shows the skills alone; ask an administrator for AI Configuration Viewer to see the connection and the models.

What does Forced do on an organization skill?

Every agent you picked lists the skill and nobody can turn it off. It needs Managed delivery and cannot be command only.

Can I promote a skill I wrote myself?

No. The message reads "You wrote or edited this skill, so another administrator has to promote it.". Another administrator has to promote it.