Saidkick is a small, opinionated tool that lets scripts, shells, and AI agents drive a real browser end-to-end — listing tabs, navigating, clicking, typing into any rich-text field, dispatching keyboard events, taking screenshots — without the overhead of a full headless automation framework. It uses a FastAPI server as a hub and a Chrome extension as the spoke; you run saidkick start, install the extension once, and then every command in every language with an HTTP client can talk to a real, logged-in Chrome session.

It's the right size for terminal-driven debugging, agent automation, and personal scripting — not quite Playwright (which needs its own browser, its own auth, its own set of tricks to look "real") and not quite a remote-control MCP (which is gated on a specific agent runtime). Saidkick lives in the middle: your browser, your session, your cookies, driven from anywhere with curl.

⚡ Features

• 🎯 Semantic locators. Target elements by what the user sees: --by-text "Send", --by-label "Password", --by-placeholder "Search…". Falls back to CSS/XPath when you need precision.
• 🧭 Scroll-into-view. saidkick scroll --tab $TAB --by-text "Chapter 3" brings an element into the viewport — essential before screenshotting something offscreen, and handy for pulling more content on infinite-scroll pages.
• 🔴 Highlight. saidkick highlight --tab $TAB --by-text "Deploy" draws a temporary red ring around an element. Use it to point the user at exactly what to click when you're guiding them — pair with screenshot and they see the ring in the image.
• 🔤 Real keyboard events. saidkick press Enter --tab $TAB dispatches a native CDP Input.dispatchKeyEvent — frameworks (Lexical, ProseMirror, React) treat it as a real keystroke, not a synthesised blob.
• 📸 Screenshots. saidkick screenshot --tab $TAB --output /tmp/shot.png via CDP Page.captureScreenshot. Optional locator clips to an element; --full-page captures beyond the viewport.
• ✍️ Rich-text input. saidkick type understands contenteditable via document.execCommand("insertText", …) — works on WhatsApp, Slack, Discord, Gmail compose, GitHub comments, Notion, and every other Lexical/ProseMirror/Quill/Slate/Draft-backed editor.
• ⏳ Wait-for-element built in. --wait-ms N on every selector-using command polls the DOM until it resolves. Default 0 preserves fail-fast behaviour.
• 🧵 Multi-browser, multi-tab. Each extension connection gets an ephemeral br-XXXX ID; commands address tabs as br-XXXX:N composites. Pipe the output of saidkick open straight into the next command.
• 🛡️ CSP bypass. Runs scripts via chrome.debugger on pages that block content-script injection.
• 🐚 Pipe-friendly CLI. One token per stdout (saidkick open prints br-XXXX:N; saidkick screenshot emits raw PNG bytes). Everything composes in bash.

🚀 Quickstart

Install

pip install saidkick

Or pull the latest from GitHub:

pip install git+https://github.com/apiad/saidkick.git

Load the extension

1. Open chrome://extensions/ in Chrome.
2. Enable Developer mode.
3. Click Load unpacked and point at src/saidkick/extension/ (inside the cloned repo, or inside your installed saidkick package — python -c "import saidkick, os; print(os.path.dirname(saidkick.__file__) + '/extension')").

The extension connects to ws://localhost:6992/ws and auto-reconnects every 5 seconds if the server comes and goes.

Start the server and drive

# Terminal 1: start the hub
$ saidkick start

# Terminal 2: list connected tabs
$ saidkick tabs
br-a1b2:12  https://example.com/  "Example Domain"  (active)
br-a1b2:15  https://docs.python.org/  "Python 3.12 Docs"

# Open a new tab, talk to it, screenshot the result
$ BR=br-a1b2
$ TAB=$(saidkick open --browser "$BR" https://example.com/)
$ saidkick text --tab "$TAB" --css "h1"
Example Domain
$ saidkick screenshot --tab "$TAB" --output /tmp/shot.png
Wrote 28934 bytes to /tmp/shot.png

🎮 Driving a chat app end-to-end

No exec, no selector archaeology — just semantic locators and a keystroke:

TAB=br-a1b2:15
saidkick click  --tab "$TAB" --by-text "Alice Chen"
saidkick type   "Hello Alice" --tab "$TAB" --by-label "Type a message"
saidkick press  Enter --tab "$TAB"
saidkick screenshot --tab "$TAB" --output /tmp/sent.png

That's WhatsApp Web, Slack, Discord, Gmail compose, or any similar app, in four lines.

🧭 Pointing the user at something

When an agent is guiding the user through an app, it often needs to say "click this button." Two primitives make that precise:

# Scroll the element into view (it may be offscreen)
saidkick scroll --tab "$TAB" --by-text "Deploy"

# Draw a temporary red ring around it (default 2s)
saidkick highlight --tab "$TAB" --by-text "Deploy"

# Screenshot so the user sees the ring in the image too
saidkick screenshot --tab "$TAB" --output /tmp/click-this.png

Good uses:

• "Click that button" — highlight + screenshot + send the image to the user.
• "The error is in this field" — highlight --color "#f59e0b" (amber) on a form field the user needs to correct.
• Pre-screenshot framing — scroll before screenshot so what you want to capture is actually in the viewport.
• Checklist walkthroughs — highlight each step as you narrate it; use --duration-ms 0 to keep the ring up until you place the next one.
• Infinite-scroll content extraction — scroll to the last visible item, wait for more to load, repeat.

scroll takes --block {center|start|end|nearest} and --behavior {auto|smooth}. highlight takes --color (any CSS color) and --duration-ms (0 = persist until page reload).

🧭 Command reference

Every selector-using command accepts the same locator options: --css, --xpath, --by-text, --by-label, --by-placeholder, --within-css, --nth, --exact, --regex, --wait-ms. Exactly one locator must be set (400 otherwise).

🐍 Python client

Everything the CLI does is also available as a library:

from saidkick.client import SaidkickClient
c = SaidkickClient()

tabs = c.list_tabs(active=True)
tab = tabs[0]["tab"]

# Search for something on DuckDuckGo
c.type(tab, "saidkick", css="input[name=q]")
c.press(tab, "Enter")

# Screenshot the results
shot = c.screenshot(tab)
import base64; open("/tmp/ddg.png", "wb").write(base64.b64decode(shot["png_base64"]))

🧱 Architecture

Hub-and-spoke. The FastAPI server is the hub; the Chrome extension (MV3) is the spoke.

┌──────────────┐      WebSocket       ┌────────────────────────────┐
│ Your CLI/    │◀───────▶ hub ────────│ Chrome MV3 extension       │
│ agent/script │    REST              │  • service worker          │
└──────────────┘                      │  • content + main-world    │
                                      │  • popup w/ reconnect      │
                                      └────────────────────────────┘

• The hub is stateless between restarts except for a circular log buffer and the set of live WebSocket connections.
• The spoke stores an ephemeral br-XXXX ID on handshake, runs content scripts in every tab on demand (with lazy injection fallback), and drives CDP via chrome.debugger for JS execution, keyboard events, screenshots, and page-load waits.
• Tabs are addressed by the composite br-XXXX:N — br-XXXX identifies the browser connection; N is Chrome's native tab.id.

The extension popup shows current connection state and a reconnect button — useful when the MV3 service worker goes idle.

📖 Docs

• User Guide — full CLI / REST / client reference.
• Design Doc — architecture, error policy, protocol details.
• Deploy Guide — server + extension setup.
• SKILL.md — how an AI agent should use saidkick.
• CHANGELOG — release history.

🤝 Why saidkick (vs. …)

• vs. Playwright / Selenium. Those spawn their own browser with a fresh profile — no cookies, no logins, no browser extensions. Saidkick drives your Chrome, logged in, with the session state you already have. Trade-off: you're automating the real thing, so destructive actions are real.
• vs. claude-in-chrome / MCP browser tools. Saidkick is self-hosted and agent-agnostic. Anything with an HTTP client can use it — shell scripts, cron jobs, arbitrary Python, any LLM runtime. Not gated on a specific agent host or credential.
• vs. raw Chrome DevTools Protocol. CDP is powerful but verbose. Saidkick wraps the patterns you actually use (locators, keyboard, screenshots, waits) behind one-line CLI commands.

🛠️ Development

git clone https://github.com/apiad/saidkick
cd saidkick
uv sync --all-groups
uv run pytest -m "not e2e"   # unit + integration
uv run saidkick start        # hub

📜 License

MIT — see LICENSE if present, otherwise standard MIT applies.