Skip to main content
Glama

Why Optima?

Ordinary browsers are made for human eyes. When an AI assistant drives one, it wastes time and tokens on ads, pop-ups and cookie walls, loses its logins, and stalls without saying anything the moment a site asks for a human.

Optima fixes the agent side of browsing, and keeps a person in the loop for the parts that need one:

Ordinary browser + AI

Optima

What the agent sees

Raw HTML or screenshots

Clean text with numbered elements: [12 button] Sign in

Cookie banners

The agent clicks around them

Declined automatically ("reject" / "only necessary")

Logins

Lost between sessions

Persistent profile: log in once

CAPTCHA / "checking your browser"

Agent gets stuck

Detected → you get a notification → you solve it → the agent continues

Passwords & cards

Risky

Always typed by you; the agent is refused these fields

Buying, posting, deleting

Happens silently

Agent must ask you in chat first

What happened?

Unknown

Daily audit log

Optima does not solve or bypass CAPTCHAs or bot protection, by design. Those checks are how sites decide who gets in. Optima makes the hand-off to you take seconds instead.

Related MCP server: Navigator MCP

How it works

 AI assistant ──MCP──▶ Optima server (Python) ──▶ engine ──▶ the web
   (Claude…)            reading, rules,             macOS: native app on system WebKit
                        hand-off, log               Windows/Linux: Edge/Chrome via Playwright
                              │
                              └──▶ you: notification + "Your turn" banner + Done button
  • macOS: a tiny native app (Swift + the WebKit that ships with macOS) that you compile on your own Mac during install. Nothing is downloaded, which helps on managed Macs where installing third-party apps is restricted.

  • Windows: uses Microsoft Edge (already on every Windows PC) or Chrome through Playwright.

  • Linux: uses Google Chrome if installed, otherwise downloads Playwright's Chromium once.

Install

You need Python 3.9+. On macOS you also need the Swift compiler (Xcode or the Command Line Tools: xcode-select --install).

macOS

git clone https://github.com/ascaniolocorotondo/optima-ai-browser.git
cd optima-ai-browser
sh install.sh

On macOS the installer also adds Optima AI Browser to ~/Applications. Double-click it to browse yourself, for example to log in to your sites: your assistant shares the same logins.

Linux

git clone https://github.com/ascaniolocorotondo/optima-ai-browser.git
cd optima-ai-browser
sh install.sh

If Chromium complains about missing libraries: sudo ~/.optima/venv/bin/python -m playwright install-deps chromium.

Windows (PowerShell)

git clone https://github.com/ascaniolocorotondo/optima-ai-browser.git
cd optima-ai-browser
powershell -ExecutionPolicy Bypass -File install.ps1

The installer copies Optima into ~/.optima, so you can move or delete the cloned folder. To update, git pull and run the installer again.

First run

The installer opens the Optima window:

  1. Choose your language. English, Italiano, Español, Français, Deutsch or Português. Banners and notifications use it; the agent always gets English.

  2. Connect your assistants. Optima detects Claude Code, the Claude Desktop app and Google Antigravity, installs its skill and registers its MCP server in each one you pick.

  3. Log in to your favourite sites in that same window, so your assistant finds you signed in.

Prefer the terminal? sh install.sh --cli (or install.ps1 --cli).

Then restart your assistant and try: "Use Optima to find the opening hours of the nearest library."

Where Optima connects

Assistant

Skill

MCP server

Claude Code

~/.claude/skills/optima/

user scope (claude mcp add-json)

Claude Desktop

upload ~/.optima/optima-skill.zip in Settings → Capabilities → Skills

claude_desktop_config.json

Google Antigravity

~/.gemini/config/skills/optima/

~/.gemini/config/mcp_config.json

Connecting other MCP clients

Any MCP client can use Optima. The server command needs no environment variables:

{
  "command": "/path/to/python",
  "args": ["~/.optima/app/src/optima_launcher.py", "mcp"]
}

On macOS /path/to/python is your python3; on Windows and Linux it is the one in ~/.optima/venv. Copy skill/optima/ into your assistant's skills folder if it supports Agent Skills.

Tools for the agent

Tool

What it does

browser_open(url)

Open a page, return it as clean text

browser_read(max_chars?)

Read the current page again

browser_click(id, confirmed?)

Click an element by its number

browser_type(id, text, submit?, confirmed?)

Type into a field (never passwords or cards)

browser_select(id, value)

Pick an option in a drop-down

browser_scroll(direction)

down, up, top, bottom

browser_navigate(action)

back, forward, reload

browser_screenshot()

PNG of the page, when visuals matter

browser_ask_human(reason)

Hand control to you and wait for Done

browser_wait(seconds)

Wait for slow pages

browser_recent_actions(limit?)

Today's audit log

The bundled skill (skill/optima/SKILL.md) only applies where Optima is turned on: if the browser_* tools are not available in a chat (for example you disabled the optima MCP server there), the assistant ignores the skill and uses its usual browser without mentioning Optima. Where Optima is on, it teaches the assistant when to use Optima (APIs and connectors first), how to act by element id, when to call you, and to always confirm purchases, messages and deletions with you.

Configuration

~/.optima/config.json (created on first run):

Key

Default

Meaning

language

chosen on first run

en, it, es, fr, de, pt

engine

auto

webkit (macOS), playwright

playwright_channel

auto

chrome, msedge, chromium

headless

false

Hide the window (Playwright only; hand-offs need a window)

identify_as_agent

true

Adds Optima/<version> (AI agent; +repo) to the user agent

dismiss_cookie_banners

true

Always the privacy-preserving choice

auto_handoff

true

Call you automatically when a human check appears

handoff_timeout_seconds

300

How long to wait for you

min_action_interval_ms

400

Gentle pace between actions

notify_desktop

true

System notification on hand-off

ntfy_topic

""

Optional phone push via ntfy: set a private topic name and subscribe to it in the ntfy app

policies.confirm_before

["purchase","send","delete"]

Actions the agent must confirm with you

policies.block

[]

Actions never allowed (same categories, plus submit)

policies.human_only_fields

true

Password and payment fields are yours only

Data lives in ~/.optima/: config.json, logs/ (one JSON-lines file per day) and, for Playwright, profile/. The macOS engine keeps cookies in the system WebKit store for the app.

Commands

python3 -m optima setup            # first-run window again (add --cli for terminal)
python3 -m optima open example.com # open Optima yourself, e.g. to log in
python3 -m optima doctor           # check the installation
python3 -m optima mcp              # the MCP server (started by assistants)

Run them with PYTHONPATH=~/.optima/app/src (or from the repo's src folder).

Principles

  • Humans decide. CAPTCHAs, logins, passwords, payments and one-time codes are always done by you.

  • Honest. Optima identifies itself as an AI agent and paces its requests.

  • Private. No telemetry. Cookie banners get "reject". Nothing leaves your machine except the pages you visit and, if you set one, your ntfy notification.

  • Accountable. Every action is logged locally.

Roadmap

  • Web Bot Auth signed requests, so sites that welcome verified agents can let Optima in without a check

  • Multiple tabs and downloads

  • Per-site rules (e.g. read-only on banking sites)

  • Packaged releases

Contributing

Issues and pull requests are welcome. Keep the principles above: features that help an agent evade CAPTCHAs, fingerprinting or bot detection won't be merged.

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      B
      maintenance
      A semantic browser runtime for AI agents that replaces raw HTML with structured data and dynamic, page-specific tools. It features built-in site memory and automated bot detection bypass to enable efficient, self-healing web automation.
      1
      -
    • A
      license
      B
      quality
      C
      maintenance
      Enables LLM agents to automate web browsers with token-efficient perception, stealth anti-detection, captcha solving, network interception, and isolated persistent sessions.
      51
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to browse and interact with web pages through structured, token-efficient representations instead of full accessibility-tree dumps, returning compact page orientations, targeted element queries, and detail on demand. Agents can navigate, click, type, fill forms, manage tabs and sessions, and inspect console or network activity at a fraction of the usual context cost.
      4 npm
      MIT