Tejas AI Agent

A Claude Code-like AI coding assistant running inside VS Code. It can read, write, and search files, run terminal commands, and chat with vision support — all from a sidebar panel powered by Anthropic's Claude.

Features

• Agentic tool use — AI autonomously reads, writes, searches files, and runs terminal commands
• Streaming responses — real-time character-by-character output
• Markdown rendering — code blocks, bold, headings render properly
• Active file context — currently open file is automatically shared with the AI
• Image / vision support — attach screenshots or images via the 📎 button
• @file mention — type @filename for autocomplete; file content is automatically included
• Tool activity log — sidebar shows what the AI is doing in real time
• Permission system — dangerous actions (file write, terminal commands) require Allow/Deny confirmation
• Persistent history — chat history survives VS Code restarts
• Secure API key storage — key is stored in OS-level encrypted SecretStorage, no .env required

Setup

1. Clone the repository

git clone https://github.com/Tejash8299/tejas-ai-agent.git
cd tejas-ai-agent

2. Install dependencies

npm install

3. Open in VS Code

code .

4. Run the extension

Press F5 — an Extension Development Host window will open with the extension loaded.

The Tejas AI icon will appear in the Activity Bar on the left.

API Key Setup

The extension requires your own Anthropic API key to work.

Get your API key:

1. Go to https://console.anthropic.com
2. Sign up or log in
3. Navigate to API Keys → Create Key
4. Copy the key (starts with sk-ant-...)

New accounts get free credits to get started. The extension uses Claude Haiku by default which is very cost-efficient.

The key is stored securely in VS Code's encrypted SecretStorage (OS-level protected). It is never written to disk in plain text and never shared.

There are 3 ways to set the key:

Option 1 — Auto prompt (first-time users)

If no key exists in SecretStorage and no .env file is present, the extension automatically shows an input box on startup:

Enter your Anthropic API Key (sk-ant-...)

Enter your key and press Enter — it is saved to SecretStorage immediately.

Option 2 — Command Palette

Ctrl+Shift+P → "Tejas AI: Set API Key"

Enter your key in the password field. Use this anytime to update or change the key.

Option 3 — .env file (development / migration)

Create a .env file in the project root:

ANTHROPIC_API_KEY=sk-ant-xxxxxxxxxxxxxxxx

On first startup, the extension reads the key from .env and automatically migrates it to SecretStorage. After migration, the .env file is no longer needed.

Note: The .env file is listed in .gitignore and will never be committed to Git.

How Tools Work

The AI has access to 5 tools. Each tool use is shown as a blue chip in the sidebar.

Dangerous tool approval flow

AI wants to use write_file or run_command
              |
              v
Yellow confirmation card appears in sidebar:
"⚠️ Allow this action?"
Shows tool name + file path or command
              |
       -------+-------
       |               |
  [✓ Allow]        [✗ Deny]
       |               |
  Tool executes    AI receives "denied"
                   and suggests an alternative

Workspace boundary protection

The AI can only access files inside the workspace folder. Any attempt to read, write, or list files outside the workspace is blocked:

Error: Access denied: "../../../etc/hosts" is outside the workspace

@File Mention

Type @ in the chat input to trigger a file autocomplete dropdown. Select a file and its full content will be appended to your message before sending to Claude.

Example:

Explain what @src/services/openai.ts does

Project Structure

tejas-agent/
├── src/
│   ├── extension.ts              # Entry point — SecretStorage, key migration, command registration
│   ├── services/
│   │   └── openai.ts             # Anthropic API client, agentic loop, streaming, tool execution
│   ├── sidebar/
│   │   └── SidebarProvider.ts    # Webview UI, chat rendering, history, permission system
│   └── tools/
│       └── index.ts              # Tool definitions, executeTool() dispatcher, safePath()
├── .env                          # Optional — gitignored, auto-migrated to SecretStorage
├── package.json
└── README.md

Key Files Explained

extension.ts

• getOrSetApiKey() — checks SecretStorage → falls back to .env → prompts user if neither exists
• Registers the tejas-agent.setApiKey command
• Passes the resolved API key to SidebarProvider

services/openai.ts

• runAgentLoop() — the main agentic while loop
• Streaming via messages.stream() and stream.on('text', ...)
• Handles stop_reason === 'tool_use' — executes tools and feeds results back to Claude
• DANGEROUS_TOOLS list — calls onConfirm() callback before executing sensitive tools

sidebar/SidebarProvider.ts

• Inline Webview HTML/CSS/JS (no external files)
• pendingConfirms Map — Promise-based approval system for dangerous tools
• workspaceState — persists agentHistory and displayHistory across restarts
• @mention regex parser — reads mentioned files and injects content into the message

tools/index.ts

• toolDefinitions — Anthropic Tool[] schema passed to the API
• executeTool() — switch-case dispatcher for all 5 tools
• safePath() — resolves and validates paths to prevent directory traversal attacks

Release Notes

0.0.3

• Secure API key storage via VS Code SecretStorage (auto-migration from .env)
• Workspace path traversal prevention (safePath())
• Tejas AI: Set API Key command

0.0.2

• Image upload and vision support
• Tool activity log
• @file mention autocomplete
• Permission system (Allow/Deny for dangerous tools)
• Improved input UI

0.0.1

• Initial release
• Agentic tool use (read, write, list, search, run command)
• Streaming responses and markdown rendering
• Active file context and persistent chat history