Quickstart · Five minutes

From a blank machine to a green agent run.

You will install the desktop app, pull a 4-GB local model, and watch the agent log into GitHub on its own. Total time: about five minutes. No API key required.

  1. 1

    Download Tanvrit Automator

    Pick the installer for your platform. Open the download page. macOS DMG is universal (Apple Silicon + Intel). Windows and Linux builds are signed.

    # macOS
open ~/Downloads/Tanvrit-Automator-1.0.0.dmg

# Windows
msiexec /i Tanvrit-Automator-1.0.0.msi

# Linux
sudo dpkg -i tanvrit-automator_1.0.0_amd64.deb
  2. 2

    Install Ollama and pull a planner model

    The default planner is qwen2.5-coder — about 4 GB. Pull it once and it stays cached on disk.

    # Install Ollama (one-time)
curl -L https://ollama.com/install.sh | sh

# Pull the planner model (one-time, ~4 GB)
ollama pull qwen2.5-coder

# Optional: pull the vision model (skip for fast first run)
ollama pull qwen2.5-vl

    On Windows, download the Ollama installer from ollama.com/download.

  3. 3

    Open Tanvrit Automator

    Launch the app. On first run it auto-downloads the Playwright browser binaries (~ 200 MB; cached afterwards) and creates the local SQLite database at ~/.automator/automator.db.

    You should see three tabs: Projects, Bench, Settings. The Settings tab confirms Ollama is reachable at http://localhost:11434.

  4. 4

    Run the sample flow

    From the menu bar choose Run → Sample: GitHub login flow. This is a pre-baked scenario that:

    • Launches Chromium
    • Navigates to github.com/login
    • Enters a sample username (configured in the scenario)
    • Stops at the password prompt — we do not store your password
    • Records every observation, action, and verification result
    # Or from the CLI:
automator run --sample github-login --headed

# Headless mode for CI:
automator run --sample github-login --headless
  5. 5

    Inspect the trajectory

    After the run completes, the right-side panel shows the trajectory: every Perceive → Plan → Execute → Verify → Record cycle, with screenshots, the LLM prompt, the chosen action, and the Verifier's before/after PageState diff.

    Click any step to scrub through. Click Export to save the trajectory as JSONL — useful for SFT training or sharing a regression case with the Tanvrit team.

    # Trajectory file location
~/.automator/automator.db
~/.automator/screenshots/<run-id>/

# Export to JSONL
automator export --run-id <run-id> --out trajectory.jsonl

Done. What next?