Integrations

Local-first by default. Cloud when you ask for it.

Everything below works in the free tier. Cloud providers fire only when you explicitly set the relevant API key environment variable. The default code path makes zero outbound calls to any third party except Ollama at localhost:11434.

Browser engine

Playwright 1.52

Stable

Drives Chromium, Firefox, and WebKit. PlaywrightManager.launch() auto-downloads matching browser binaries on first run.

DESKTOP / MOBILE / TABLET device profiles supported. Headless and headed both work; headed is preferred during development for visual debugging.

Local LLM provider

Ollama

Stable

The default LLM backend. Runs entirely on your machine. No API key, no rate limit, no token charges.

qwen2.5-coder for planning (~4 GB), nomic-embed-text for RAG embeddings (~270 MB), qwen2.5-vl for vision (~5 GB). Set OLLAMA_URL to point at a remote Ollama if you want a shared model server.

Cloud LLM provider

Anthropic Claude

Opt-in

Cloud fallback when your local model can't fit a huge prompt. Set AUTOMATOR_LLM_PROVIDER=anthropic and ANTHROPIC_API_KEY.

Default model is claude-sonnet-4-6. Tanvrit charges nothing extra; you pay Anthropic directly. The default code path is local — cloud calls fire only when you explicitly opt in.

OpenAI

Opt-in

GPT-4o and o-series models supported via the OpenAI client. Set OPENAI_API_KEY and AUTOMATOR_LLM_PROVIDER=openai.

Google Gemini

Opt-in

Gemini 1.5 Pro and Flash supported. Set GEMINI_API_KEY and AUTOMATOR_LLM_PROVIDER=gemini.

Groq

Opt-in

Fastest inference for Llama / Mixtral / Qwen via the Groq Cloud API. Set GROQ_API_KEY.

DeepSeek

Opt-in

DeepSeek-V3 and Coder via the DeepSeek API. Set DEEPSEEK_API_KEY.

Mistral

Opt-in

Mistral Large and Codestral via the Mistral API. Set MISTRAL_API_KEY.

Tool exposition

MCP — Model Context Protocol

Stable

Tanvrit Automator can expose its 31 ToolDispatcher actions as MCP tools, callable from any MCP-compatible client.

Wire it into Claude Desktop, Cursor, or Zed and your LLM coding assistant gains direct browser-driving capability. Both stdio and HTTP transports are supported. See /help/mcp.

Local storage

SQLDelight + SQLite

Stable

All trajectories, bench flows, and projects are persisted to ~/.automator/automator.db.

SQLDelight 2.1.0 generates type-safe Kotlin from SQL schemas in commonMain/sqldelight/. Tables: flows, trajectories, trajectory_steps, bench_flows, bench_runs, projects, action_steps, ui_patterns.

Vision fallback

Tesseract OCR

Stable

Last-resort OCR for pages where Qwen2.5-VL captioning misses on-screen text.

Install with `brew install tesseract` (macOS) or `apt install tesseract-ocr` (Ubuntu). Language packs are separate; install the ones your target apps use.

CI integration

GitHub Actions

Phase 2

A reusable Action that runs Tanvrit Automator headless against a deployed preview URL — bench gates for your app's CI.

Currently you wire it manually with a JVM runner; the published action is on the roadmap. The bench runner CLI is functional today.

Environment-variable cheat sheet

# Default — local Ollama, zero cloud calls
AUTOMATOR_LLM_PROVIDER=ollama
OLLAMA_URL=http://localhost:11434
OLLAMA_MODEL=qwen2.5-coder

# Anthropic Claude (opt-in)
AUTOMATOR_LLM_PROVIDER=anthropic
ANTHROPIC_API_KEY=sk-ant-...
ANTHROPIC_MODEL=claude-sonnet-4-6

# OpenAI / Gemini / Groq / DeepSeek / Mistral
AUTOMATOR_LLM_PROVIDER=openai      # or gemini, groq, deepseek, mistral
OPENAI_API_KEY=sk-...

# Smart autofill (opt-in, step-0 form fill)
AGENT_SMART_AUTOFILL=on