vedant-codex 0.3.0

Local AI-agent for science & technology

Vedant Codex

A lightweight AI-powered coding assistant CLI inspired by tools like Codex and Cursor.

Vedant Codex allows developers to interact with an AI agent directly from the terminal to analyze projects, edit files safely, search codebases, and automate development tasks.

It integrates with Ollama and tool-enabled models to provide a local or hybrid AI coding workflow.

---

Features

• AI coding assistant inside your terminal
• Safe file editing using unified diff patches
• File system access restricted to allowlisted directories
• Native .docx Word document creation and editing
• Native .pptx PowerPoint creation and editing
• Native .xlsx Excel workbook creation and editing
• Native image inspection and OCR-style readability
• Browser automation for real webpage interaction
• Built-in tools:
  • create files
  • read files
  • read Word documents (.docx)
  • read PowerPoint decks (.pptx)
  • read Excel workbooks (.xlsx)
  • inspect images (.png, .jpg, .jpeg, .webp, .gif, .bmp, .tiff)
  • browser automation (open, inspect, click, type, select, wait, read, scroll, back, keypress, screenshot)
  • search files
  • list project structure
  • apply patches
  • write Word documents (.docx)
  • write PowerPoint decks (.pptx)
  • write Excel workbooks (.xlsx)
  • run commands
  • open files in browser
  • start development servers
  • web search
• Supports tool-calling capable models
• Session logging
• Rolling token-based quota window
• Cross-platform support

---

Installation

Recommended (pipx)

Install using pipx for isolated CLI environments.

1. pip install pipx
2. pipx ensurepath
3. pipx install vedant-codex

Alternative (pip)

1. pip install vedant-codex

Run

1. vedant

Prerequisites

-Python (Python 3.10 or newer is recommended.)

-Ollama (Vedant Codex requires Ollama to run AI models.)

Install Ollama:

https://ollama.com/download

Initialize and download ollama.exe.

Usage

Start the assistant: vedant

Example prompts:

Create a login.html page with modern glassmorphism UI Search for TODO comments in this repository Start a Python development server List files in this project

Configuration

Environment variables can customize behavior.

Example .env (default): MODEL=qwen3-coder:480b-cloud VISION_MODEL=qwen3-vl:235b-cloud OLLAMA_HOST=http://127.0.0.1:11434 ALLOWLIST_ROOTS={{add C:/Users/path/to/workingspace or implicitly give permission to model to use any path you want.}} DAILY_MAX_TOKENS=200000 BROWSER_MODE= BROWSER_CHANNEL= BROWSER_HEADLESS=0 BROWSER_USE_SAVED_PROFILE=0 BROWSER_USER_DATA_DIR= BROWSER_PROFILE_DIR= BROWSER_CDP_URL= COMMERCE_BROWSER_HEADLESS=0 COMMERCE_BROWSER_USER_DATA_DIR= COMMERCE_BROWSER_PROFILE_DIR=Default COMMERCE_DEFAULT_PAYMENT=cod

Browser automation requires Playwright browsers after install: python -m playwright install chromium

You can also drive an installed Microsoft Edge or Chrome browser channel. Useful browser settings:

• BROWSER_MODE=background for a dedicated persistent automation profile
• BROWSER_MODE=personal_edge to attach to a live personal Edge session over CDP
• BROWSER_CHANNEL=msedge to prefer Microsoft Edge
• BROWSER_HEADLESS=1 to keep browser automation in the background
• BROWSER_USE_SAVED_PROFILE=1 to try reusing the browser's regular saved profile when supported
• BROWSER_PROFILE_DIR=Default to target a specific browser profile inside the user data directory
• BROWSER_USER_DATA_DIR=... to point at a dedicated browser user-data folder if you do not want to reuse the default profile
• BROWSER_CDP_URL=http://127.0.0.1:9222 for the personal Edge attach endpoint

Commerce/browser-shopping defaults:

• mode="commerce" opens a dedicated Microsoft Edge AI-agent profile instead of your personal profile
• COMMERCE_BROWSER_USER_DATA_DIR=app\storage\browser-profiles\msedge-ai-agent to pin the dedicated Edge agent profile location
• COMMERCE_BROWSER_PROFILE_DIR=Default to keep a single reusable profile directory inside that agent profile
• COMMERCE_DEFAULT_PAYMENT=cod to make cash on delivery the default checkout preference
• On the first commerce run, sign in manually once on the sites you use; the dedicated profile keeps those logins for later sessions

For reliable background shopping, prefer a dedicated automation profile over your everyday Edge Default profile. A common Windows setup is: BROWSER_MODE=background BROWSER_CHANNEL=msedge BROWSER_HEADLESS=1 BROWSER_USER_DATA_DIR=app\storage\browser-profiles\msedge-shopping

Use BROWSER_USE_SAVED_PROFILE=1 only if you explicitly want to try the normal Edge profile and are okay with launch failures when Edge is already open or the profile is locked.

For personal Edge mode, start Edge yourself with remote debugging enabled and then let the agent attach after you explicitly approve it. Example on Windows: msedge.exe --remote-debugging-port=9222

With that running, the agent can use mode="personal_edge" and your real logged-in Edge session.

From this repo on Windows cmd, the simplest launcher is: start-personal-edge-debug.cmd

Then verify the debug endpoint: curl http://127.0.0.1:9222/json/version

For real shopping and account flows, the browser agent can navigate, search, fill non-sensitive forms, add to cart, and reach checkout across sites like Amazon, Flipkart, and Myntra. It will stop before final payment or order placement unless the user explicitly confirms the risky step.

For commerce and travel flows in v0.3.0, the agent can also keep a dedicated Edge AI-agent profile for sites like amazon.in, flipkart.com, myntra.com, irctc.co.in, cleartrip.com, and makemytrip.com. Use browser_set_checkout_context plus browser_review_checkout so the agent tracks COD/card preference, reviews detected payment options, and asks for a final confirmation before any order, booking, or subscription activation.

Security

Vedant Codex is designed with safety in mind.

File system access is restricted to allowlisted directories

File edits must use reviewable diff patches

Command execution can be restricted with allowlists

Token quotas prevent runaway automation

Development

Clone the repository: git clone https://github.com/Probro-2009/local-codex Then: cd local-codex Then: pip install -e . Then: vedant

Roadmap

Planned improvements:

persistent memory

improved tool orchestration

better project indexing

model auto-detection

improved Windows / Termux compatibility

plugin system

Contributing

Pull requests are welcome.

If you find a bug or want a new feature, email at [e-mail].