Ghost Browser — Agentic control of Dhruv's real Chrome

You give it a goal in plain English. It drives his actual logged-in Chrome like a human — scrolling, clicking, typing, reading — and stops before any irreversible public action.

Runtime lives at: Social-Media-Agent-1.0/browser-automation/ Engine: browser-use (0.13.x) + Playwright CDP attach + real Google Chrome Brain: OpenAI gpt-5.4-mini (actions) + gpt-5.4 (accurate page reading), keys in that folder's .env

Why this beats normal automation

• Un-blockable: drives your REAL Chrome with your REAL logged-in sessions + residential IP — same trick Manus uses. Not a headless bot, so bot-detection has nothing to catch.
• Persistent: Chrome opens ONCE and stays open. No reopening tabs per action.
• Human-paced: small scrolls, brief pauses, deliberate steps — fast where safe, slow where blocking is a risk.
• LLM-driven: you state the goal; it decides the clicks. Survives LinkedIn/any site renaming its buttons (no brittle selectors).
• Safe on the brand account: never clicks a final Submit / Post / Send / Apply-confirm, or reacts to >3 posts per run. It fills everything and STOPS for Dhruv's click. Override with --allow-submit.

How to run it (two steps)

Step 1 — attach to a profile (once per session). Quits Chrome, clean-copies the login, relaunches on the debug port. His everyday Chrome is untouched (works on a copy).

cd "Social-Media-Agent-1.0/browser-automation"
./.venv/bin/python ghost_launch.py aiv     # or: onsite | msbc | dhruv | agi | default
# waits for "✅ READY"

Step 2 — command it in plain English (repeat as many times as you want; Chrome stays open).

./.venv/bin/python ghost_agent.py "open Jobs, search Solution Consultant remote India, list the top 5 with company + location"
./.venv/bin/python ghost_agent.py "scroll my feed and draft a thoughtful comment on the top AI post — do NOT post it"
./.venv/bin/python ghost_agent.py "open this job <url>, fill the application with my details, stop before Submit"

Profile map (name → Chrome profile)

Safety model (do not weaken without Dhruv's explicit OK)

• Reading, scrolling, navigating, searching, opening posts, and typing into comment/message/application fields = automated.
• Final public/irreversible actions (Post, Send, Submit, Publish, Connect, >3 reactions/run) = Dhruv clicks, unless the command includes --allow-submit.
• This protects the brand account from a ToS flag mid-job-hunt.

Files

Notes / gotchas

• Chrome 136+ blocks the debug port on the normal profile dir — that's why ghost_launch.py uses a clean copy under .chrome-cdp-<name>/ (gitignored).
• Set a different brain with GHOST_MODEL / GHOST_READ_MODEL env vars. Euri keys are present for cost-sensitive product runs.
• The engine has a built-in LLM-judge that grades each run — watch for its verdict to catch misreads.
• Requires the runtime built: python3 -m venv .venv && .venv/bin/pip install browser-use playwright && .venv/bin/playwright install chromium.

Alternative: Claude for Chrome (official, complementary)

For interactive "Claude, do this in my browser" moments, the official Anthropic Claude for Chrome extension (Max plan) is the safest zero-code option — records/repeats workflows, schedules tasks, uses live sessions. Ghost Browser is the scriptable, ownable version for automation Dhruv controls.