Configuring AI

Agents, response modes, system prompts, and ability toggles

Every workspace has AI agents that can participate in conversations. The built-in HUD agent is set up automatically, and you can create additional agents — each with their own personality, capabilities, and rules for when they respond. Configuration happens at two levels: the agent itself (who it is, what it can do) and the channel (when it responds).

Agents

An agent is an AI identity in your workspace. Manage them in Server Settings > Agents.

Built-in agents

Built-in agents use start.chat’s AI pipeline. Each one supports:

• Name and avatar — give the agent its own identity in conversations
• System prompt — customize the agent’s personality, instructions, and behavior. This replaces the default HUD persona entirely.
• Model tier — xxs, xs, sm, md, lg, or xl, depending on your speed/quality needs
• Abilities — toggle which capabilities the agent has (search, chatDocs, zeroQueries, blocks, appActions, flows, skills, memory, workspace)

HUD is created automatically as a built-in agent. You can customize its system prompt and settings just like any other agent.

External agents

External agents connect via webhook endpoint and respond through the start.chat API. They get an API key for authentication and can declare capabilities (code, shell, browser, files). Create one by choosing “External Agent” when adding a new agent. For a complete example using a real agent framework, see OpenClaw.

Channel configuration

Each channel independently controls which agents respond and how. Open the channel’s settings and go to the Agents tab.

Adding agents to a channel

Click the add button to see available agents. Selecting an agent adds it with the default response mode of Auto-reply (thread).

Response modes

Each agent-channel pair has its own response mode:

Setting the mode to Off removes the agent from the channel entirely.

Auto-reply (thread) is the default and works well for support or Q&A channels — every question automatically gets its own thread, keeping the main channel clean.

Mentions only is best for channels where you want AI available but not intrusive. Users @mention the agent when they need help.

Note that @mentioning an agent always triggers a response regardless of the configured response mode.

Multiple agents

You can configure several agents in the same channel with different response modes. A code-review agent might auto-reply in threads in #engineering while a general-purpose agent only responds when @mentioned.

Custom system prompts

The system prompt is the most powerful configuration lever. It replaces the default HUD persona entirely. Some examples:

A code reviewer:

You are a code reviewer for our React codebase. Focus on:
- Type safety and proper TypeScript usage
- Component composition patterns
- Performance implications

Be direct. Skip pleasantries. If the code is good, say so briefly.

A support agent:

You are a support agent for Acme Corp. Be friendly and thorough.
Always check the knowledge base before answering.
If you can't find the answer, say so and suggest opening a ticket
in #support-escalation.

System prompts are set per-agent in Server Settings > Agents. If you need different behavior in different channels, create separate agents with different prompts and assign them to the appropriate channels.

Ability toggles

Each ability can be enabled or disabled per-agent:

Disabling an ability removes both its tools and related prompt instructions. An agent with only Search and Skills enabled won’t attempt to run queries or generate blocks — it won’t even know those capabilities exist.

File reading is always active and does not appear in the toggle list.

Advanced configuration

For fine-grained control, click the Advanced button in the agent’s abilities section. This opens a JSON editor where you can set:

{
  "modelTier": "lg",
  "temperature": 0.3,
  "maxTokens": 32768,
  "disabledAbilities": ["blocks", "flows"]
}

If enabledAbilities is set, only those abilities activate (minus any in disabledAbilities). If only disabledAbilities is set, all defaults remain active except those listed.