AI Overview
What the AI can do, how it decides when to respond, and how to configure it
Start has a built-in AI assistant called HUD. You can talk to it in a dedicated heads-up display, @mention it in any channel, or configure it to auto-reply. Under the hood, it has a composable set of abilities that let it chat, search, remember, take actions, and more.
What it can do
HUD has eight abilities, all enabled by default:
- Search — look up curated documentation (library) or fuzzy/vector search across your workspace’s chats and data (lookup). Library returns higher-quality curated results; lookup searches everything.
- Memory — remember things you tell it across conversations. Say “remember that…” and it stores the fact for future use. See AI Memory for details.
- Queries — read and modify workspace data. The AI can run predefined queries to fetch channels, threads, messages, and other records. It can also propose mutations (inserts, updates, deletes) that require your approval before executing.
- App actions — interact with installed apps. Each app can expose actions the AI can discover and run on your behalf. For example, a CRM app might expose a “search contacts” action.
- Flows — create and manage automations. Flows link app actions together and can be triggered manually, on a schedule, or in response to events. The AI can generate flow definitions right in a message.
- Blocks — generate inline UI widgets (Stat, Chart, Status, List, Table) that render directly in chat. It knows about available block types, their props, and how to compose them.
- Skills — read workspace-specific documentation added by admins. Skills provide curated context the AI can load on demand when answering questions.
- Platform docs — look up documentation about Start itself when you ask how things work.
The AI can also view file attachments (text, CSV, JSON, code, markdown, PDF) when they’re shared in conversation.
Response modes
Each channel can configure when the AI responds. This is set per-agent in channel settings:
| Mode | Behavior |
|---|---|
| Auto-reply (thread) | Responds to every message, creating a thread first if not already in one. Good for keeping the main channel clean. |
| Auto-reply (inline) | Responds to every message directly in the channel. |
| Threads only | Only responds when you’re already inside a thread. |
| @mention only | Only responds when you @mention it. This is the default for most channels. |
| Off | Never responds in this channel. |
You can always @mention HUD regardless of the channel’s response mode. The dedicated HUD page (the heads-up display) always responds.
Model tiers
The AI uses a tiered model system with seven tiers. Server admins set the default tier, and HUD selects models within that tier based on task complexity:
| Tier | Use case | Primary models |
|---|---|---|
| xxs | Quick operations | Gemini Flash Lite, Claude Haiku |
| xs | Simple tasks | Gemini Flash, Claude Haiku |
| sm | Tool-calling/strategy | Gemini Flash (high thinking), Claude Haiku |
| md | General conversation | Claude Haiku (extended thinking), Gemini Flash |
| lg | Detailed responses | Claude Sonnet, Claude Haiku |
| xl | Complex reasoning | Claude Opus, Claude Sonnet |
The default tier is md. Admins can override the tier per-channel or per-server through the AI configuration.
Configuring the AI
AI behavior can be customized at the channel or server level:
- Model tier — choose between xxs, xs, sm, md, lg, and xl
- Enabled/disabled abilities — toggle specific abilities on or off. For example, disable block generation in a non-technical channel.
- System prompt — override the default persona. The AI defaults to “HUD, a helpful AI assistant” but you can replace this with custom instructions for specific channels or agents.
- Temperature — control response randomness (0-2, model default if unset)
- Max tokens — cap response length (default 16384)
Mutation approval
When the AI wants to modify data (send a message, create a channel, update a record), it outputs a mutation block that renders with an Accept button. The mutation only executes after you approve it, and your permissions apply — the AI can’t do anything you couldn’t do yourself.
HUD mode vs channel mode
In HUD mode (the dedicated AI interface), conversations are private between you and the AI. You can create multiple chat threads in a sidebar, similar to ChatGPT.
In channel mode, the AI participates alongside other users. It sees the channel context, server name, and recent conversation history. How much history it sees depends on context: in threads it uses the last 10 messages; outside threads it walks back to find the most recent @mention or up to 4 non-bot messages.