Getting started with VritraAI

This guide takes you from zero to a finished agent task: install, keys, model, first prompt, and what to do next.

What is VritraAI?

VritraAI is a terminal-native AI coding agent. You open a workspace folder, configure a provider, and describe what you want built or fixed. The agent plans with tools, shows diffs for approval, runs shell/verify when needed, and calls finish_task when done.

Version
1.0.2-cli
Python
3.9+
License
MIT
Platforms
Linux · macOS · Termux
Author
Alex Butler / Vritra Security Organization

Installation

Install from PyPI (recommended):

pip install vritraai

Requires Python 3.9+ and network access for model APIs. Full options (Termux, editable source install): see Installation.

1. Launch

Open the folder you want the agent to work in. Paths, writes, and shell stay inside that workspace.

vritraai ~/projects/my-app

On first open, choose a permission mode:

  • ask - preview writes and shell; approve each time (safest default).
  • auto - trust this workspace (fewer prompts).
  • safe - stricter / skip some prompts carefully - see Approvals.

The choice is saved per folder under .vritraai/trust.json.

2. Configure API keys

/setup

/setup walks only the providers you have not configured yet. Paste a key (masked input). Empty / skip skips that provider.

Supported providers: Gemini, OpenRouter, Groq, Mistral, NVIDIA.

Keys are stored in ~/.config-vritrasecz/vritraai/config.json - outside any git repo. You only need to set them once.

A free Gemini key or OpenRouter free models is enough to start.

3. Pick a model

/model

Interactive ↑/↓ picker over models for providers that already have keys. Good free-tier defaults:

  • gemini-flash-latest
  • openrouter/free or any OpenRouter id ending in :free
  • Groq open-weight models as failover

Tip: Selection persists. If a model dies mid-task, failover can switch automatically when other keys/models are available.

4. Your first agent task

No slash prefix means the coding agent. Be concrete about paths and acceptance criteria.

Build a landing page in site/ with index.html, styles.css, and app.js - real content, not stubs.

What you should see:

  • Live narration (Calling model… Create… $ shell…)
  • Diff / write previews - approve, always this session, or deny
  • Optional verify steps (compile / pytest)
  • A finish summary when the agent calls finish_task
vritraai › Scaffold a FastAPI service with /health, pytest, and a short README

· workspace /home/you/demo
· Calling model …
· Create  app/main.py
· Create  tests/test_health.py
· $ python3 -m pytest -q
Done - health route green, README written.

Slash commands vs agent

InputGoes to
Plain English (no /)Coding agent loop
/review, /project, /doc, …Focused slash utilities
/ then Tab / typingAutocomplete menu

Keyboard & UI

  • Tab - slash autocomplete
  • ↑ / ↓ - history / picker navigation
  • Ctrl+C - cancel current operation
  • /clear - clear the screen
  • /exit - quit

File changes use divider-style previews (not noisy boxes). Streaming narration can be disabled with VRITRA_AGENT_STREAM=0.

Next steps