AI Abilities
What the AI can do, and how to extend it
Most chat platforms give you a bot that can answer questions. Start gives you an AI that can take real actions — search your knowledge base, remember things across conversations, run queries against your data, execute app actions, generate code, and orchestrate multi-step flows. Each of these capabilities is called an ability.
What is an ability?
An ability is a self-contained package of tools and context that gets composed into the AI’s prompt at request time. When you message HUD (or any built-in agent), the system assembles the response by:
- Loading context — memories about you, workspace knowledge
- Activating abilities — each one contributes tools the AI can call and instructions on when to use them
- Building the system prompt — persona, context, ability instructions, tool guidelines, then conversation history
- Streaming the response — the AI can call any tool from any active ability mid-response, up to 100 tool steps per message
Abilities are lazy. The AI gets lightweight descriptions of what each tool does, then calls the ones it needs. This keeps the prompt manageable and responses fast.
Built-in abilities
All abilities are enabled by default. Each can be individually toggled per-agent in Server Settings > Agents.
Search
Two tools for finding information:
library searches curated documentation. When you ask about a framework, library, or concept, the AI searches the docs index and extracts the relevant sections. If one document is comprehensive enough, it uses that alone. Otherwise it pulls from the top results and notes what else is available.
lookup searches across messages, data, and content within your workspace. Results are scoped to the current server and filtered through a secondary pass to remove duplicates and stale content. Results include citation links so the AI’s response can reference the original messages or data records.
You don’t need to think about which tool to use — the AI picks automatically.
Memory
Say “remember that our deploy target is us-east-1” and the AI stores it. Next conversation, that context is automatically part of the system prompt without the AI announcing “I remember that…”.
Memories are scoped to you within the current workspace. Up to 20 memories are loaded per conversation, each truncated to 200 characters in the prompt, so keep them concise. See Memory for the full system.
Data
The AI can query and modify your workspace data through predefined, permission-filtered queries. It can look up table schemas and run queries to answer questions like “how many messages were sent this week?” or “show me the latest threads in #engineering.”
For modifications, the AI proposes changes that you must explicitly approve. When the AI wants to create a channel, send a message, or update data, it renders the proposed change with an Accept button. Nothing executes until you click it, and your own permissions apply — the AI can’t do anything you couldn’t do yourself.
App Actions
Installed apps can expose actions that the AI calls on your behalf. The AI discovers available actions on demand — it sees which apps are installed, then fetches the action list for a specific app when it seems relevant.
If you have the GitHub app installed, asking “check the status of my open PRs” will trigger the AI to call the GitHub app’s fetch action. The AI can also generate entire apps inline — describe what you need and it can write the code for a block that renders in your channel.
Flows
Flows let the AI create and manage automated pipelines that chain app actions and AI steps. In a normal channel, the AI can write flow definitions. In the dedicated Flows channel, it’s aware of the specific flow being managed and can query its current state.
Blocks
The AI generates inline UI widgets — Stat, Chart, Status, List, Table — that render directly in chat. It loads available block types and their props, then composes them to display data, metrics, or status information.
Skills
Skills are documents created by workspace admins that the AI reads on demand. Style guides, API references, onboarding docs, internal procedures — anything you want the AI to know about but don’t want cluttering every prompt.
The AI lists available skills and reads specific ones when a question seems relevant. If your admin added a “deployment procedures” skill, asking “how do I deploy to staging?” pulls it into context automatically.
Platform Docs
The AI has access to documentation about start.chat itself. Questions like “how do threads work?” or “what are channel groups?” pull up the relevant platform docs.
Workspace
The AI can manage your personal workspace settings and channel-level preferences:
- Settings — change your theme, notification preferences, locale, and quiet hours. These are applied directly since they only affect you.
- Channel notifications — control per-channel notification levels (all, mentions, none) and set temporary mutes. You can target individual channels or all channels in a sidebar group at once.
- Agent response mode — tell HUD to stop responding in a channel, switch to mentions-only, or change how it auto-replies. If you say “stop responding” or “be quiet”, the AI will offer to adjust its response mode.
- Reminders — HUD can set up reminders for you using scheduled flows. Ask something like “remind me every Wednesday at 9am about standup” and it creates a recurring notification.
File Reading
This ability activates automatically when files are attached to the conversation. The AI sees a summary of all attachments with their types and sizes, and can read the contents of text-based files (code, CSV, JSON, markdown, PDFs) with pagination for large files. File reading is always active and cannot be disabled.
How apps provide abilities
Apps extend the AI through two mechanisms:
Actions. Any installed app can declare actions in its metadata. These become callable through the App Actions ability. The AI discovers them on demand and executes them with a message argument.
Persistent context. Apps can write data records with
aiVisibility: 'context' that are automatically injected into every AI
conversation. The memory ability uses this pattern internally — when you say
“remember X”, it writes a data record that future conversations pick up. Any app
can do the same to surface persistent context to the AI. See
Memory for details.
System prompt assembly
When the AI responds, the system prompt is assembled in this order:
- Persona — the agent’s identity. HUD uses a default persona; custom agents use their configured system prompt
- Runtime context — which server, channel, and thread you’re in
- Memories and knowledge — your saved memories and workspace-level knowledge, loaded with a 500ms timeout
- Ability instructions — one section per enabled ability, listing available tools and guidance on when to use them
- Tool guidelines — priorities for tool usage
- Conversation history — the recent messages in the current thread
The ordering is intentional. Persona and context establish identity. Memories sit between identity and capabilities because they’re about who you are, not what tools exist. Ability instructions are the dynamic part. Conversation history comes last as the direct input being responded to.