Tappi: Your Browser on Autopilot - Without the Token Tax
A local, token-efficient AI agent that controls your real browser. No APIs. No bans. No screenshots-as-input madness.
The Problem Nobody Talks About
Every AI browser agent today does one of two things: send a full screenshot to the LLM, or dump the entire DOM tree. Sometimes both.
Screenshots mean the LLM is doing vision work - squinting at pixels, guessing coordinates, praying it clicks the right button. A single interaction can burn thousands of tokens on image analysis alone.
DOM dumps are even worse. A modern webpage's DOM is enormous - tens of thousands of tokens of nested divs, classes, aria attributes, and scripts. The LLM has to reason through all of it just to figure out which element to interact with. That's not automation. That's an LLM doing a reading comprehension exam on every single click.
OpenAI's Operator, Anthropic's Computer Use, even most open-source browser agents - they all pay this tax. The LLM spends more time figuring out what's on the page than actually doing your task. Scale that to any real workflow and you're hemorrhaging tokens while the agent overthinks where to click.
What Tappi Does Differently
Tappi doesn't send screenshots. It indexes every interactive element on the page into a compact numbered list:
[0] (link) Skip to content
[1] (button) Toggle navigation
[2] (link) Homepage → https://github.com/
[3] (button) Platform
[4] (link) GitHub Copilot - Write better code with AI
[5] (link) GitHub Spark - Build and deploy intelligent apps
The LLM sees this and says click 4. Done. No pixel parsing. No coordinate guessing. No hallucinated CSS selectors.
The result: 10x fewer tokens per interaction. The LLM reasons less and acts faster. You can visibly see the difference - where other agents pause and deliberate on every click, Tappi fires through actions at machine speed.
Here's a real session - the agent browsing GitHub Trending, showing every tool call:
The numbered element list is what the LLM actually sees - not a 1MB screenshot.
And here's what Tappi's browser was looking at:
Why Local Matters
"Just use an API." "Just use a scraping service."
Sure - if you want to:
- Get rate-limited or banned from platforms
- Pay per request for data you could get for free
- Lose access to anything behind a login
Tappi runs on your machine, in your browser, with your saved sessions. Log into Gmail, Twitter, LinkedIn, Instagram once - and Tappi can work with all of them forever. No re-authentication. No CAPTCHAs. No proxy rotation.
Because it's doing the same thing you'd do manually - just faster and in the background - there's zero ban risk. You're not scraping. You're browsing.
And your data never leaves your machine. No cloud. No third-party servers.
Who Is This For?
Anyone who uses a browser. That's not a cop-out - it's the point.
- Social media managers - Schedule posts, reply to comments, monitor engagement
- Researchers - Scrape data, compile reports, track competitors. Output straight to PDF or Excel
- Founders - Automate outreach, monitor mentions, pull analytics
- Developers - Test flows, scrape docs, automate repetitive dev tasks
- Anyone drowning in tabs - Hand off the busywork. Keep the thinking.
Sandboxed by Design
Most AI agent platforms have access to your filesystem, your shell, your everything. That's powerful - but it's a lot of trust.
Tappi is deliberately limited:
- One browser. One workspace directory. That's the blast radius.
- No filesystem access beyond the folder you define
- No shell access unless you explicitly enable it
If your work is browser + files (and for most people, it is), Tappi gives you everything you need without exposing your system.
Features at a Glance
| Feature | Details |
|---|---|
| 🌐 Browser Control | Navigate, click, type, scroll, screenshot, tab management |
| 📁 File Management | Read, write, move, copy - sandboxed to workspace |
| Read existing PDFs, create new ones from HTML | |
| 📊 Spreadsheets | Read/write CSV and Excel (.xlsx) |
| ⏰ Cron Jobs | Schedule recurring tasks with cron expressions |
| 💻 Shell | Optional - run commands within workspace |
| 🖥️ Web UI | Chat interface with live tool-call visibility |
| 🤖 Multi-Provider | OpenRouter, Anthropic, OpenAI, Claude Max (OAuth), Bedrock, Azure, Vertex |
| 🔍 Shadow DOM | Pierces Shadow DOM (Reddit, GitHub, etc.) |
| 👤 Profiles | Multiple browser profiles - work, personal, social |
Getting Started
1. Install
pip install tappi[agent]
Requirements: Python 3.10+, Chrome or Chromium. Works on Linux, macOS, Windows.
2. Setup
bpy setup
The wizard walks you through provider, API key, workspace, and browser profile:
Settings are also configurable via the Web UI.
Pro tip: If you have a Claude Pro/Max subscription ($20-200/mo), use your OAuth token - the same one Claude Code uses. No per-call API charges.
3. Launch Your Browser
bpy launch
First launch opens a fresh Chrome window. Log into the sites you want to automate - Gmail, GitHub, social media. Those sessions persist for all future launches.
Create separate profiles for work, personal, and social media.
4. Give It a Task
CLI - one-shot:
bpy agent "Go to Hacker News and summarize the top 5 posts into a PDF"
CLI - interactive:
bpy agent
Web UI:
bpy serve
Here's what a real session looks like. I asked: "Go to github.com/trending and tell me the top 3 trending repos today"
The agent launched the browser, navigated to the page, indexed all elements, extracted the text, and reported back:
Each tool call is visible in real-time as the agent works.
Final result with structured analysis.
5. Schedule Recurring Tasks
You: Every morning at 9 AM, check trending repos on GitHub
and save the top 10 to trending.csv
Agent: Done. Created job "GitHub Trends" - runs daily at 09:00.
Jobs persist and run automatically when the server is active.
The Architecture
You (CLI or Web UI)
↓
┌─────────────┐
│ LLM Agent │ ← Sees 6 tools as JSON schemas
└──────┬──────┘
│
┌──────┴──────┐
│ Tool Calls │
├─────────────┤
│ 🌐 Browser │ → CDP → Your Chrome (with all your sessions)
│ 📁 Files │ → Sandboxed workspace directory
│ 📄 PDF │ → Read/create PDFs
│ 📊 Sheets │ → CSV/Excel
│ 💻 Shell │ → Optional, workspace-only
│ ⏰ Cron │ → Scheduled recurring tasks
└─────────────┘
No middleware. No cloud. No screenshots. Just structured element data flowing between your browser and your LLM.
CLI Quick Reference
# Agent
bpy setup # Configure provider, workspace, browser
bpy agent [message] # Chat (interactive or one-shot)
bpy serve # Start the Web UI
# Browser
bpy launch # Start Chrome
bpy open <url> # Navigate
bpy elements # List clickable elements (numbered)
bpy click <n> # Click element by number
bpy type <n> <text> # Type into element
bpy text # Extract visible text
bpy screenshot [path] # Save screenshot
bpy tabs # List open tabs
Links
- GitHub: github.com/shaihazher/tappi
- PyPI: pypi.org/project/tappi
pip install tappi[agent]
bpy setup
bpy agent "Check my Gmail for unread emails and summarize them"
Three commands. That's it. It opens your Chrome, uses your saved login, reads your inbox, and reports back.
Built by Azeruddin Sheik. Tappi started as an internal tool for browser automation - then it got efficient enough to stand on its own.
Top comments (0)