Extend Bolt
Skills and agents.
Teach the AI in Bolt a prompt or a tool, or add an agent of your own. Personal, if your organization allows it.
Skills.
Settings → Skills. A skill you make is yours. Your organization can turn off adding your own skills, and on some plans skills are not available.
| Format | What it does | Tier |
|---|---|---|
prompt | Instructions added to the conversation when the skill is used. | 1 |
tool | A tool the AI can call. It makes one HTTP request to an endpoint you describe. | 3 |
agent_task | A scoped agent run with its own tools and limits. | 3 |
| Field | Meaning |
|---|---|
name, description | What it is. |
format | prompt, tool or agent_task. |
content | The prompt text, or the tool definition in YAML. |
tags, trigger_word | Find it and start it by a word. |
name: ticket_status
description: Look up a support ticket by number
parameters:
type: object
properties:
id: { type: string, description: "Ticket number" }
required: [id]
instructions: Use when the user asks about a ticket number.
endpoint:
url: "https://support.example.com/api/tickets/lookup"
method: POST
body_template: '{"id": "{{id}}"}'
The URL is fixed; the AI's arguments go into body_template as {{name}}.
Agents.
Settings → Agents. Use one by typing @
and its name in a chat, or with one of its trigger words. Your
organization can turn off adding your own agents.
| Field | Meaning |
|---|---|
name, description, instructions | Who the agent is and how it works. |
backend | Where it runs: Bolt's own loop, or a remote endpoint (OpenAI-compatible, custom HTTP, or A2A). |
tool_allowlist, tool_blocklist, mcp_allowlist | Which tools and MCP servers it may use. |
max_iterations, reasoning_strategy | How long it may work. |
triggers | Keywords that send a request to it. |
background_capable, icon, tags | Run in the background; display. |
An agent with tools is tier 3: it can act through every tool it is allowed. Keep its tool list short.