This page is the reference for the Advanced features list under Settings > Developer: what each feature does, the risk you accept when you turn it on, and what it needs. Turning features on and off, and accepting their risks, is covered in Advanced features; the Developer settings page covers the area the list sits in.
Which features the list offers depends on the app you run and how your computer is set up.
Oda can take screenshots of your screen and move the mouse, click and type in any app on this computer.
Risk: A web page, file, email or code comment the agent reads can hold hidden instructions that make it control your desktop: open apps, send messages or change files outside your workspace. Anything on your screen, including passwords and private messages, can reach the model.
Applies immediately. Works best with a model that reads images. Computer use covers the feature in the app.
Oda opens its own Chrome or Edge window and can browse, click, fill in forms and read pages.
Risk: Pages the agent visits can hold instructions that make it submit forms, download files or send what it read to other sites. It starts with a clean browser profile without your logins; connecting it to your own browser would hand it your signed-in sessions and cookies.
Applies to new chats. Needs Google Chrome or Microsoft Edge. Browser tools covers the browser in a chat.
HTML files open in the side panel with their scripts running, so interactive pages, charts and games work.
Risk: Scripts in an HTML file the agent wrote or downloaded run inside the app. A malicious page can show fake prompts or try to reach services on your computer or network.
Applies immediately. While it is off, HTML previews show the page without running its scripts. Preview images and HTML covers the preview.
When you ask for it, Oda can start a long unattended loop that runs a fresh agent round after round until the task reports itself done.
Risk: A loop can run for a long time without you and spend many tokens. "Done" is the agent's own report, not an independent check, so review the result before you rely on it.
Applies immediately.
A real PowerShell or shell terminal opens in the right side panel.
Risk: Commands typed there run with your full rights on this computer and are not checked by the chat's permission mode.
Applies immediately. The terminal covers the terminal tab and what you can do with it.
Adds an Auto permission mode: before each tool call a model reviews the action and blocks risky ones instead of asking you.
Risk: Experimental. The reviewer is a model and can block good actions or allow unsafe ones. Auto mode has Full access rights, so anything the review allows runs without asking you.
Applies immediately. Needs a DataMind account with Full access.
Oda shows the Plugins page, where you install, turn on and turn off plugins, including plugins from outside DataMind.
Risk: A plugin runs inside the app with your full rights and can read your chats, files and saved credentials. Install only plugins you trust.
Applies immediately.
Microsoft sign-in pages in the Sidebar browser sign you in with the work account Windows is signed in with, the way Edge does and Chrome does under its CloudAPAuthEnabled policy. A Conditional Access policy that only admits compliant devices then sees this computer.
Risk: The agent drives the same browser tabs, so it can open any site your work account reaches, such as mail, files and admin portals, without a password prompt. Pages it visits can hold instructions that make it act there as you.
Applies to the next sign-in page. Needs a Windows computer joined to Microsoft Entra ID or with a work account added.
Browser use works in a browser window Oda opens itself: Oda browses, clicks, fills in forms and reads pages. Computer use works across your desktop: Oda can take screenshots, move the mouse, click and type in any app. They are separate switches with separate risks; Browser tools and Computer use cover each.
Each feature's row says when. Most apply immediately: Computer use, Interactive HTML preview, Ralph autonomous loop, Terminal tab, Auto review permission mode and Plugins page. Browser use applies to new chats, and Windows work account sign-in applies to the next sign-in page.
No. Each feature asks once for itself, the first time you turn it on, and its row then records who accepted that feature's risks and when (Risks accepted by {by} on {date}). Accepting one feature does not accept another.
No. "Done" is the agent's own report, not an independent check, so review the result before you rely on it. A loop can also run for a long time without you and spend many tokens.
No. Commands typed in the terminal tab run with your full rights on this computer and are not checked by the chat's permission mode. See The terminal.
A Windows computer joined to Microsoft Entra ID, or with a work account added. It applies to the next sign-in page, where Microsoft sign-in pages in the Sidebar browser use the work account Windows is signed in with.
A plugin runs inside the app with your full rights and can read your chats, files and saved credentials. Install only plugins you trust.
Yes. Anything on your screen, including passwords and private messages, can reach the model. Content Oda reads can also carry hidden instructions that make it control your desktop, so turn the feature on only when you need it, and watch what Oda does.
Before each tool call, a model reviews the action and blocks risky ones instead of asking you. The reviewer is a model, so it can block good actions or allow unsafe ones; anything it allows runs without asking you, because Auto mode has Full access rights.