Skip to content

Capabilities ​

The Capabilities switches under the composer's + decide which tools Halo's AI can use in a conversation, such as Web Control, AI Terminal and Operate Halo.

Capabilities are optional tool packs built into Halo. Turn one on and the AI gains a new kind of skill in that conversation, such as working in web pages, opening a terminal or managing Halo for you. When the AI needs a capability it asks you to turn it on, but it can't turn one on itself.

Curious what an AI desktop agent can do on your own computer? See AI desktop agent.

Where to find them ​

Click + in the composer (tooltip "Add files, context and capabilities"). The panel has three groups: Add, Context and Capabilities. The switches are in the Capabilities group. Click one to turn it on or off.

The composer + panel, with the Capabilities group listing Web Control, Team Collaboration, AI Terminal, Text Recognition and Operate Halo

  • While the AI is replying, the + panel shows only the Capabilities group, and the tooltip reads "Capabilities".
  • The switches belong to each conversation. Another conversation can have a different set.
  • In a chat with a Digital Human, the + panel has no capability switches. A Digital Human's tools are set in its own settings; see How this differs from a Digital Human's switches.
  • Halo opened through Remote Access can flip these switches too.

A capability that's off takes up one summary line in the conversation. A capability that's on brings its tools and usage notes into the conversation, so every message uses a few more tokens. Turn on only what the task needs.

The five switches ​

SwitchDescription in the appOn first useDetails
Web ControlAI works in your browser, with your signed-in sessionsOnAI Browser
Team CollaborationFor bigger tasks, AI assembles a team that works in parallelOnDigital Human teams
AI TerminalInteractive terminal you share with AI: SSH, other CLIs, long tasksOffAI Terminal
Text RecognitionReads text in images on this device, no tokens usedOffFiles and images
Operate HaloAI operates Halo for you: spaces, digital humans, knowledge bases, settingsOffLet the AI operate Halo

"On first use" means the state before you have ever flipped a switch by hand. What happens after that is under Defaults and when changes apply.

Web Control ​

With Web Control on, the AI can open pages, click, fill in forms, take screenshots and download files in Halo's built-in AI Browser. Sites you've signed in to in the AI Browser stay signed in for the AI. Pages open on the computer that runs Halo.

Team Collaboration ​

With Team Collaboration on, the AI can assemble a temporary team for a bigger task, act as its coordinator and let the members work in parallel. When you ask, it can keep that team, or hand the work to a team you've already saved. How teams split, hand over and report back is on Digital Human teams.

AI Terminal ​

macOS and Windows only

With AI Terminal on, the AI can open a persistent interactive terminal that you and the AI both see and type into. It suits signing in to a server over SSH, driving other command-line tools and long-running jobs. The Linux version has no such switch. See AI Terminal.

Text Recognition ​

With Text Recognition on, the AI can read the text in images on your computer, in English and Simplified Chinese. Images aren't uploaded, and recognition itself uses no tokens. It extracts text only; it doesn't describe what an image shows.

When you send images to a model that can't see images, Halo turns Text Recognition on for that conversation automatically, and it works for that very message. Details: Files and images.

Operate Halo ​

With Operate Halo on, the AI can use Halo's own API to manage workspaces, conversations, Digital Humans, teams, knowledge bases, message channels and some settings, and it can switch models. General settings such as theme and language stay out of its reach. What it can and can't do is on Let the AI operate Halo.

Caution

Operate Halo is off on first use. Once it's on, the AI can act on any workspace, not just the current one, and nothing asks you to confirm first. Only for a few actions that can't be undone does Halo tell the AI to check with you first. Turn it off when you're done. No Halo tool call asks for per-call confirmation; see Tool permissions.

Defaults and when changes apply ​

SituationWhat happens
You've never flipped a switchNew conversations start with Web Control and Team Collaboration on, the rest off
You've flipped switches beforeNew conversations use the set you last flipped by hand
You reopen an older conversationIt keeps its own set
You flip a switch while the AI is replyingThe current reply finishes as is; the change applies from your next message
You send images to a model that can't see imagesHalo turns Text Recognition on for that conversation, effective for that message
You upgraded from 2.xTeam Collaboration is added to your saved set once; if you turn it off afterwards, it stays off

When the AI asks for a capability ​

When a task needs a capability that's off, the AI neither pretends it can do the task nor turns the capability on. Instead it:

  1. Opens the + panel and briefly highlights that switch. The panel doesn't take keyboard focus, so whatever you're typing isn't interrupted.
  2. Asks you in one line to turn it on, then stops and waits.

Turn the switch on and send another message, and the AI can use the capability. For example, you say:

SSH into the test server and check the latest error logs.

If AI Terminal is off, AI Terminal is highlighted in the + panel and the AI asks you to turn it on. Turn it on, reply "done", and it carries on.

The AI can ask this way even in the middle of a reply. Operate Halo doesn't change this: turning capabilities on or off isn't part of the API that Halo opens to the AI, so only you can flip them.

Tools that need no switch ​

These tools are in every ordinary conversation. They aren't in + › Capabilities and need nothing turned on:

ToolWhat the AI can doDetails
Web searchSearch the web and return results, without relying on the model's own search—
Halo docs lookupRead Halo's official docs to answer "how do I do this in Halo"Docs for AI tools
Digital Human managementCreate, edit, schedule, run now, pause and delete Digital Humans; install or uninstall skillsCreate a Digital Human
Cross-conversation toolsRead other conversations in the same workspace, hand work to one, or ask it a question and wait for the replyConversations that work together

On top of these, the AI has the basic tools of its AI engine, such as reading and writing files and running commands.

Turn some built-in tools off in Settings ​

Desktop only

Web search and Halo docs lookup have no switch. Digital Human management and the cross-conversation tools can be turned off in Settings › Advanced › Built-in MCP Extensions. In every UI language, this panel's descriptions and every row name except Digital Humans are shown in English:

Switch (as shown)What it controlsDefault
Digital HumansDigital Human management toolsOn
Cross-Conversation InteropMaster switch for the cross-conversation tools: list, read and message other conversations in the same workspaceOn
Cross-Conversation SendingLets the AI send to other conversations; off means read only. Applies only while the row above is onOn

The same panel's Plan Mode, Git Worktree, Scheduled Tasks and Notebook Editing are features of the AI engine itself and are off by default. Changes take effect in a new session or after you restart Halo. Each row is described in the Settings reference.

How this differs from a Digital Human's switches ​

+ › Capabilities covers your own conversations. Each Digital Human has its own set of switches under Digital Humans › pick one › Capabilities and settings › Model & Capabilities:

  • The switches are AI Browser, AI Terminal, Operate Halo, Conversation Collaboration, Email and IM Push.
  • All are on by default except Operate Halo and Conversation Collaboration.
  • The page notes "Always available: web search, memory". Text recognition is always available to Digital Humans too.
  • These switches decide which tools that Digital Human can use in its chats and runs. They don't affect the switches in your composer, and yours don't affect them.

What each switch means and its limits: Capabilities and settings.

How this differs from the Capability library ​

The Capability library lives on the Digital Humans page, split into Skills and MCP connections. It holds the skills you install and the MCP servers you connect. Once installed, they're available to the conversations and Digital Humans in their scope, with nothing to turn on under + › Capabilities.

+ › Capabilities has only the five built-in switches above. You can't add new ones to it.

I want to…Use
Let the AI use a browser or terminal, or manage Halo, this timeTurn on the matching switch in + › Capabilities
Give the AI a fixed way of working, such as a weekly-report routineSkills
Connect the AI to an outside system's toolsMCP servers
Decide which tools a Digital Human can useThat Digital Human's Capabilities and settings
Stop the AI managing Digital Humans or reading other conversations in ordinary chatsSettings › Advanced › Built-in MCP Extensions

Next steps ​

基于 MIT 协议开源 · Open source under the MIT license