Agent Overview

Agents are standardized AI workflows that complete specific, repeatable tasks by combining prompts, rules, and tools.

Agents are customized AI workflows composed of prompts, rules, and tools (such as MCP and other integrations) used to execute specific, reusable tasks. They can be hosted in the Agent Hub and run via a web interface, or created and used locally within the IDE.

Agent Components

Component	Description	Example
Model	The large language model that drives agent reasoning and output generation.	Qwen3
Rules	Specify consistent standards or behavioral guidelines that the agent must follow when responding, ensuring reliability of usage results.	"When summarizing work, must include relevant Issues."
Prompts	Instructions that trigger the agent's core behavior. When invoked, user input is appended to this prompt.	Example prompt: "Summarize current work progress, including status, blockers, and next steps. Use Markdown format, language concise and professional."
Tools/MCP	Extend external capabilities (databases, APIs, CLI)	GitHub, PostHog, Supabase

Agent Types

VJSP AI Studio supports three types of agents, distinguished by management method:

• Official Agents (used via reference)

• Hub Custom Agents

• Local Agents

Agent Structure and Component Configuration

All local agents are defined via config.yaml, located in the .vjsp/agents/ folder in the project root directory.

Basic Structure

# This is an example configuration file
name: Config    # Required
version: 1.0.0  # Required
schema: v1  # Required

# Define which models can be used
models:   # Model configuration
  - name: 
    provider: openai
    model: 
    apiKey: YOUR_OPENAI_API_KEY_HERE

Core Component Configuration Explanation

Models

Define the language model that drives the agent.

models:
  - name: 
    provider: openai
    model: 
    apiKey: YOUR_OPENAI_API_KEY_HERE
    roles:
      - chat
      - edit
      - apply
    capabilities: 
      - tool_use
    requestOptions:
      headers:
        vjsp-api-key: your_api_key

• roles: Specify model purposes (chat/edit/autocomplete/apply)

• capabilities: Enable advanced features (such as tool_use for calling MCP)

Rules

Enforce AI to follow coding standards.

Format as follows:

---
description: A description of your rule
---

Your rule content

Prompts

Quickly trigger tasks via /commands.

---
name: New prompt
description: New prompt
invokable: true
---

Please write a thorough suite of unit tests for this code, making sure to cover all relevant edge cases

MCP Tools (mcpServers)

Connect to external systems and databases.

name: New MCP server
version: 0.0.1
schema: v1
mcpServers:
  - name: New MCP server
    command: npx
    args:
      - -y
      - <your-mcp-server>
    env: {}

Custom Local Agent

Operation Steps:

Step 1. Open VJSP Plugin in IDE

Ensure you are logged in and the VJSP AI Studio plugin is enabled.

Step 2. Create Configuration File

• Open the Agent Management panel in the IDE

• Click Settings → Configuration → Add Configuration

• The system will automatically create the .vjsp/config.yaml file in the project root directory

Step 3. Fill in Basic Information

Edit config.yaml to define agent metadata and model configuration. Example as follows:

name: TestCaseGenerator
version: 1.0.0
schema: v1

models:
  - name: Qwen3
    provider: openai
    model: Qwen3
    apiBase: https://xxx.xx.cn/v1
    capabilities:
    - tool_use    # Enable tool calls (required for MCP)
    roles:
    - chat        # Support chat mode
    - edit        # Support edit mode
    - apply       # Support applying generated results
    requestOptions:
        headers:
            vjsp-api-key: b4xxxxxx

📌 Field Description

capabilities: [tool_use] is a prerequisite for using MCP tools

roles determine in which IDE modes this model is available

Step 4. Configure Core Components (Key Step)

• User Settings: Plugin global settings, set font size, formatting, code block auto-wrap, etc.

  

• Configure Models: Specify LLMs used for different modes

  

• Configure Rules: Enforce coding standards or behavioral constraints

<!-- 换图  -->

• Configure Prompts: Define task templates triggered by /commands

• Tool Configuration: Integrate external capabilities like databases, APIs, CLI

  

Step 5. Enable Agent

• Save all configuration files

• Click 🔄 Reload Agents in the IDE

• Switch to the newly created local agent at the top of the chat window

• Start using (supports Chat / Edit / Agent modes, etc.)

Important Notes

• 🔒 Scope Limitation: Local agents are only valid for the current project and are not automatically synchronized to the web side.

• 🔄 Activation Mechanism: After modifying configuration, you must manually reload for changes to take effect.

• 🧪 Experimental Features: Advanced capabilities like multi-turn tool calls, complex workflows may require enabling "Experimental Agent Features" in plugin settings.

• Reuse Recommendation: If you need to use across multiple projects, create web-side agents in the Agent Hub.