automatiq 0.1.2

Record browser sessions and reverse-engineer them into automation scripts.

Your activity, into automation.

AutomatiQ

[!Note] Alpha 🡒 Things will break and change. Read VISION.md to understand why Automatiq exists and where it's headed.

AutomatiQ watches you browse, then an AI agent reverse-engineers your session into a standalone Python automation/extraction script; no manual inspection needed.

What it does

1. Record ⭢ Opens Chrome, captures your browsing (screen video, network requests, user actions). Press Ctrl+C when you're done.
2. Compile ⭢ Vision AI analyses video clips around each action; network requests are decoded, deduplicated, and structured into a workspace dump.
3. Agent ⭢ An LLM investigator reads the workspace, experiments in a sandboxed IPython environment, and iteratively produces a working script.

Quick start

pip install automatiq

Set your API key (any litellm-supported model):

GEMINI_API_KEY=your-key-here

Run:

automatiq run https://example.com

That's it. Browse the site, press Ctrl+C, and the agent takes over.

Keyboard shortcuts

Phase	Key	Action
Recording	Ctrl+C	Stop recording and save session
Compilation	Esc	Skip AI analysis for remaining segments
Compilation	y / n	Confirm or deny the skip prompt
Agent	q	Quit the agent session
Agent	Esc	Cancel current LLM call or code execution

Ctrl+C force-quits at any phase.

CLI options

Flag	Description
--model MODEL	LiteLLM model string for the agent
--recorder-model MODEL	Vision model for video-clip analysis
--base-url URL	Custom OpenAI-compatible API endpoint
--max-steps N	Maximum agent loop iterations (default: 60)
--sandbox-timeout SEC	Seconds per IPython cell (default: 60)
--output-dir PATH	Root directory for all output (default: ./output)
--no-banner	Skip the startup animation
--verbose	Show detailed diagnostic output
-V, --version	Show version
-h, --help	Show help message

How it works

• Browser capture — Chrome is launched with CDP instrumentation. Every network request, response body, cookie, and user interaction (clicks, typing, navigation) is recorded with timestamps.
• Vision analysis — The recording is split into per-action video clips. A vision LLM watches each clip and produces structured annotations (what was clicked, what changed, whether the action succeeded).
• Sandboxed agent — The investigator runs Python code in an isolated IPython worker process. It can read the captured data, test hypotheses against the live site, and build the final script incrementally, with guardrails against loops and repetition.

Configuration

On first run, AutomatiQ creates ~/.automatiq/config.toml with commented defaults. Edit it to override models, timeouts, recording settings, etc.

[models]
agent    = "gemini/gemini-3-flash-preview"
recorder = "gemini/gemini-3.1-flash-lite-preview"
# base_url = "http://localhost:11434/v1"   # Ollama / LM Studio / vLLM

[agent]
max_steps       = 60
sandbox_timeout = 60

[recording]
fps                   = 3
segment_pad           = 2
merge_gap_threshold   = 1.5
max_frames_per_prompt = 8

Priority: CLI flag > ~/.automatiq/config.toml > built-in defaults.

Step-by-step usage

automatiq record https://example.com   # just record
automatiq agent                         # build automation script from last recording

Install from source

AutomatiQ is managed using uv.

git clone https://github.com/StoneSteel27/AutomatiQ.git
cd AutomatiQ
uv sync
uv run automatiq run https://example.com

Dev setup

Development dependencies (pytest, ruff, pre-commit, etc.) are installed automatically when you run uv sync. To set up the git hooks:

uv sync
uv run pre-commit install

Run tests and benchmarks:

uv run pytest

This ensures ruff, build, twine, pytest, and pre-commit hooks (lint + format on every commit) are properly configured in your isolated environment.

Requirements

• Python 3.11+
• A supported LLM API key (Gemini, OpenAI, OpenRouter, or any OpenAI-compatible endpoint via --base-url)

License

MIT