Example
const config: AIAssistant.AgentConfiguration = {
systemPromptTemplate: 'You are an expert legal document assistant.',
skills: [
{
name: 'contract-review',
description: 'Analyze contracts for key terms and risks',
content: '1. Identify all parties\n2. Extract key dates\n3. Highlight liability clauses'
}
],
toolApproval: {
defaults: { default: 'allow', write: 'ask' }
}
}
Properties
mcps
Model Context Protocol (MCP) HTTP servers to connect the AI Agent to.
Each entry connects the agent to one HTTP/HTTPS MCP server. The server exposes tools and resources via the Model Context Protocol; the agent can invoke those tools as part of completing a user request.
Only the HTTP transport is supported. To use a stdio-only MCP
server, expose it behind an HTTP gateway and configure that gateway
here.
Authentication, and any other headers required for the MCP server,
are provided via the headers field on each server entry. Two common
patterns:
Example
// Bearer token (RFC 6750):
{
transport: 'http',
name: 'analytics',
url: 'https://mcp.example.com/',
headers: { 'Authorization': 'Bearer ' + accessToken },
}
// API key in a custom header (vendor convention):
{
transport: 'http',
name: 'crm',
url: 'https://api.example.com/mcp',
headers: { 'X-API-Key': 'sk-...' },
}
Default Value
[]seedMessages
Seed messages for shaping agent behavior across all interactions.
Messages injected before the conversation history on every model invocation. Useful for few-shot prompting, additional constraints, and behavioral priming that should persist across all conversation turns.
Important: Seed messages are NOT persisted to the conversation history and are injected fresh on each request.
Common use cases:
- Few-shot learning: Include human/ai example pairs
- Additional instructions: Constraints that complement systemPromptTemplate
- Behavioral priming: Consistent patterns to reinforce across all responses
| Property | Description |
|---|---|
| content | The text content of the message. Use clear, concise language. For few-shot examples, make responses representative of the desired behavior. |
| type | The sender type for this message.
|
Example
seedMessages: [
{
type: 'human',
content: 'Analyze the indemnification clause in this contract.'
},
{
type: 'ai',
content: 'I found the indemnification clause on page 3, section 5.2...'
}
]
skills
Domain-specific skills for the AI Agent.
Define specialized knowledge, instructions, or workflows that the agent can access and apply when relevant to user requests. Skills allow you to inject domain expertise without modifying the system prompt.
The agent decides when to use each skill based on the skill description and the user's request. Use descriptive names and clear instructions.
| Property | Description |
|---|---|
| content | Detailed knowledge and instructions for this skill. The actual content the agent uses when applying this skill. Include step-by-step instructions, examples, or knowledge the agent should follow. Maximum 50,000 characters. |
| description | Description of when and how to use this skill. Helps the agent decide when to apply this knowledge based on user requests. Be specific about the skill's purpose and ideal use cases. Keep it under 500 characters. |
| name | Unique identifier for this skill. Used internally by the agent to reference and invoke this skill. Should be lowercase with hyphens (e.g., 'contract-analysis', 'data-extraction'). Minimum 1 character, no spaces. |
Example
skills: [
{
name: 'contract-review',
description: 'Analyze contracts for key terms and identify potential risks',
content: `When reviewing a contract:
1. Identify all parties and signatories
2. Extract critical dates (effective date, expiration, renewal)
3. Summarize payment terms and amounts
4. Highlight liability and indemnification clauses
5. Flag unusual or potentially risky terms`
}
]
systemPromptTemplate
Custom system prompt template for defining agent personality and behavior.
Defines the base instructions, role, and guidelines that shape how the agent processes requests and responds to users. This is the primary mechanism for customizing the agent's expertise and behavior patterns.
Keep prompts concise but descriptive. Focus on the agent's role, key principles, and important constraints.
Example
systemPromptTemplate: `You are an expert legal document analyst.
Your role is to:
- Identify and explain key contractual terms
- Highlight potential risks or unusual clauses
- Always cite the specific section numbers
- If uncertain, clearly state your limitations`
Default Value
A general-purpose document assistant prompttoolApproval
{ … }
Tool execution approval policies for security and user control.
Controls which tools can execute automatically and which require user approval. Use this to implement security policies, prevent unintended modifications, or require confirmation for sensitive operations.
Resolution order (first match wins):
- Pattern-based rules (if defined)
- Tool category defaults (read/write based on tool metadata)
- Fallback default setting
| Property | Description |
|---|---|
| defaults | Default approval settings for tools by category. These defaults apply when no rule matches. The resolution order is:
|
| rules | Pattern-based rules for fine-grained tool control. Rules are evaluated in order - the first matching pattern determines the approval mode. Place more specific patterns before general ones. Patterns use standard glob syntax where |
Example
// Basic: Allow reads, ask for writes
toolApproval: {
defaults: {
default: 'allow',
read: 'allow',
write: 'ask'
}
}
// Advanced: Pattern-based rules with defaults
toolApproval: {
defaults: { default: 'allow' },
rules: [
{ pattern: '*_delete_*', approval: 'deny' },
{ pattern: 'mcp__*_modify_*', approval: 'ask' }
]
}
userId
User identifier for rate limiting and authorization.
Provide a unique identifier for the current user to enable per-user rate limiting and usage tracking.
Example
userId: 'user12345'
Advanced configuration for AI Agent behavior and capabilities.
Customize the agent's system instructions, define domain-specific skills, control tool execution policies.
This interface enables fine-grained control over how the AI Agent processes requests, what knowledge it has access to, and which operations require user approval.