VertexAgent - Autonomous Private AI Coding Agent

Developed by VertexLabs
Copyright Zafer Kilicaslan
www.vertexlabs.de

VertexAgent is a VS Code extension that provides an autonomous AI agent with multi-provider support, automatic error correction, and persistent memory.

Features

Multi-Provider Support

Google Gemini - Direct API integration
Anthropic Claude - Claude 3.5 Sonnet support
OpenAI-compatible - llama.cpp, LiteLLM, Ollama
Ollama - Local models without API key
Custom - Any OpenAI-compatible API

Autonomous Agent Loop

Automatic testing - Code is tested after creation
Error detection - VS Code diagnostics integration
Auto-fix - Errors are fixed automatically
Iterative improvement - Loop runs until code is clean

Persistent Memory System

Design decisions - Architecture notes are stored
Context continuity - Memory persists across sessions
Automatic storage - memoryNotes in .vertex/memory.json
Auto-initialization - .vertex/memory.json is created automatically
Memory context in chat - Recent notes are injected into new chat requests

Intelligent Code Processing

Automatic file creation - Code blocks are extracted and written as files
Multi-file support - Create multiple files in one response
File/directory deletion - DELETE support for files and folders (recursive)
Relative paths - Workspace-based path resolution
Command-only mode - For folder/file requests, the agent can return pure Bash commands
Safe Bash FS application - mkdir/touch/rm -rf from Bash code blocks are applied safely inside the workspace
Hardened prompts - Security rules, testing requirements, structured logging standards

Modern UI

Minimalist design - Dark theme with VertexLabs branding
Diff visualization - Code blocks show +X -Y statistics with syntax highlighting
Collapsible code blocks - First 4 lines visible, click header to expand
Command blocks - Terminal-style UI for bash commands
Live status feedback - Spinner shows current activity
Markdown rendering - Formatted responses with lists and headings
Token tracking - Right-aligned "Used Tokens:" display

Configurable Settings

Provider selection - Dropdown with automatic configuration
Dynamic fields - UI adapts to selected provider
API key management - Secure storage in VS Code settings
Token options - Checkbox for optional authentication

Installation

Requirements

Node.js >= 18
VS Code >= 1.85
LLM provider (Gemini API key, Ollama, or llama.cpp server)

Setup

Clone the repository:

git clone https://github.com/vertexlabs/vertex-agent.git
cd vertex-agent

Install dependencies:

npm install

Compile extension:

npm run compile

Load extension in VS Code:

Press F5 to launch debug mode
Or package as .vsix: npm run package

Configuration

Set up a provider

Open settings:

Click the ⚙️ button in chat
Or: Cmd+Shift+P -> "VertexAgent: Settings"

Gemini API

Provider: Google Gemini
Server URL: https://generativelanguage.googleapis.com
Server Port: 443
API Key: Your Gemini key (from https://makersuite.google.com/app/apikey)

Ollama (Local)

Provider: Ollama (local)
Server URL: http://localhost
Server Port: 11434
API Key: Leave empty

Start Ollama:

ollama pull llama3.2
ollama serve

llama.cpp (Local)

Provider: OpenAI-compatible
Server URL: http://localhost
Server Port: 8080
Use access token: ☐ (disabled)

Start llama.cpp:

./server -m model.gguf -c 4096 --port 8080

Usage

Open Chat

Cmd+Shift+P -> "VertexAgent: Open Chat"
Or configure a shortcut

Commands

Create code:

Create a Python function that calculates Fibonacci numbers

Create multiple files:

Create a complete Python CLI tool with:
- main.py (argument parsing)
- utils.py (helper functions)
- test_main.py (unit tests)

Fix errors:

Fix the TypeScript errors in src/app.ts

Folder/file via command-only mode:

Create a folder named docs and a TODO.md file inside it

Expected agent response example:

mkdir -p docs
touch docs/TODO.md

Delete files/folders:

Delete the test_folder directory

Expected agent response example:

rm -rf test_folder

These commands are applied as safe filesystem actions in the current workspace.

Live Status

During processing you will see:

"Thinking..."
"Collecting context..."
"Processing response..."
"Writing 3 file(s)..."

Architecture

Project Structure

vertex-agent/
├── src/
│   ├── agent/
│   │   ├── loopEngine.ts        # Autonomous agent loop
│   │   ├── memoryEngine.ts      # Persistent memory
│   │   ├── errorAnalyzer.ts     # VS Code diagnostics
│   │   ├── contextBuilder.ts    # Project context
│   │   └── types.ts             # Type definitions
│   ├── ai/
│   │   ├── aiClient.ts          # LLM client
│   │   └── providerAdapter.ts   # Multi-provider support
│   ├── fs/
│   │   └── fileEditEngine.ts    # File operations
│   ├── ui/
│   │   ├── chatPanel.ts         # Chat interface
│   │   └── settingsPanel.ts     # Settings UI
│   └── extension.ts             # Extension entry point
├── .vertex/
│   └── memory.json              # Persistent memory
└── package.json

Components

LoopEngine - Orchestrates agent iterations with auto-test and auto-fix
MemoryEngine - Persists design decisions
ErrorAnalyzer - Collects VS Code diagnostics for error correction
ProviderAdapter - Converts API formats (Gemini, Claude, OpenAI)
FileEditEngine - Applies edits in the workspace
ChatPanel - Minimal UI with markdown and status feedback

Design Decisions

Provider adapter pattern - Unified interface for different LLM APIs
Automatic code extraction - Regex-based parsing of code blocks
Persistent memory - JSON-based storage in .vertex/
VS Code diagnostics integration - Uses native error detection
Minimalist UI - Focus on functionality over noise
TypeScript - Type safety for a robust extension

Development

Compile

npm run compile

Watch mode

npm run watch

Package extension

npm run package

Debugging

Press F5 in VS Code
Extension loads in a new VS Code window
Console logs appear in "Debug Console"

Testing

See TESTING.md for the full test suite.

Provider Documentation

Gemini Setup

See GEMINI_SETUP.md for detailed Gemini API integration guidance.

Supported Models

Provider	Model	Context	Cost
Gemini	gemini-flash-latest	1M tokens	Low
Claude	claude-3-5-sonnet	200k tokens	Medium
Ollama	llama3.2	128k tokens	Free
llama.cpp	Custom	Variable	Free

Known Issues

Gemini rate limits - Heavy request volume can trigger rate limits
Ollama performance - Local models are slower than cloud APIs
Memory size - Very large memory files can affect performance

Recently Fixed (v1.5.7-1.5.9)

✅ Agent interpreting refactoring requests as DELETE operations (fixed in v1.5.8)
✅ File creation requests triggering DELETE instead of code output (fixed in v1.5.7-1.5.8)
✅ mkdir only creating first directory when multiple specified (fixed in v1.5.9)

Roadmap

[x] Diff visualization in chat (v1.5.0)
[x] File/directory deletion support (v1.5.0)
[x] Command blocks with terminal UI (v1.5.0)
[x] Hardened system prompts with security rules (v1.5.0)
[ ] Streaming support for realtime responses
[ ] Git integration for automatic commits
[ ] Custom system prompts per project
[ ] Workspace-specific settings
[ ] Plugin system for custom providers

License

Proprietary
Copyright VertexLabs - Zafer Kilicaslan

The VertexLabs logo and all VertexAgent branding elements are proprietary.

Support

Website: www.vertexlabs.de
Developer: Zafer Kilicaslan