For web, Reddit and X searches, article reading, bookmark archiving and bounded timeline reading, use Start Browser Hands.cmd on Windows (Start X Reader.cmd is an alias). With Python 3.11+ and Chrome installed, you can also install requirements-x-reader.txt and run python x_reader_app.py. This separate interface opens at http://127.0.0.1:8772. It does not need an LLM API key. Research saves rendered source text, links and verbatim excerpts locally; see research workflows and limits.
Sign in manually in the reader's dedicated Chrome profile, close that window normally, then start a scan. Posts are saved locally as they appear and exported as Markdown/JSONL. Topic files use keyword matching. Daily timeline reading is off until configured and runs only while the reader is open. Browser capture is partial, and restarting scans from the top is not exact cursor continuation. See X Reader setup and limitations.
To archive one whole X account (all posts, replies, quotes, links and media, a week at a time and resumable) into an Obsidian vault with a connection graph, and optionally Notion, double-click Archive X Profile.cmd. See Archive a whole X account.
X prohibits website scripting; slow reading does not remove account risk. The generic browser agent and its social digest remain separate from this reader.
See the Codex additions and handoff for the full change inventory, verification, unfinished Show more support, runtime setup and future work.
An autonomous web operations suite with a developer-grade Bento Box console. Execute complex browser tasks across authenticated web applications (LinkedIn, Google Cloud, AWS, GitHub), maintain persistent sessions safely, monitor feeds with zero-hallucination verification, sync task execution logs & extracted data to a self-updating Notion Kanban database, and access your agent securely from your workstation, LAN, or mobile device.
The primary interface provides 1-click starter prompts, workflow templates, inference model controls, multi-turn conversational thread memory, real-time live execution telemetry, and persistent profile authentication.
Toggle Turbo Mode (Sandbox) to instantly bypass the 30-second about:blank LLM discovery turn with zero code risk. Click ⚡ Optimize Prompt to transform loose, conversational requests into high-precision, multi-phase execution blueprints with exact Boolean queries and extraction schemas.
The fast compiler analyzes unstructured requests (e.g. searching LinkedIn for multi-company recruiters and firing connection requests), automatically synthesizing Boolean search strings, field extraction tables, rate-limit guardrails, and structured markdown output schemas.
Connect an existing Notion database or auto-provision a new Kanban board with 1 click. Tasks, real-time status transitions (Queued → In Progress → Needs Human → Completed), chronological CDP steps, and fetched content are automatically populated into Notion.
Read-only, human-paced feed scrolling with DeepSeek v4.1 Flash. The agent deduplicates posts via SHA-1 fingerprints, reads post links in background CDP tabs, and verifies extracted summaries against verbatim DOM body text.
Persistent turn history with real-time state caching, status chips (DONE, FAILED, STOPPED), ⚡ TURBO and ✨ OPTIMIZED execution tags, timing metrics, and non-jittery expandable execution summaries.
- Instant Pre-Navigation (Turn 0): Deterministically parses destination URLs from user instructions (
youtube.com,linkedin.com,console.cloud.google.com) and opens them immediately in the initial tab action. Eliminates the 20–40s blank-page DOM screenshot & evaluation delay. - CDP Ad & Tracker Suppression: Uses Chrome DevTools Protocol
Network.setBlockedURLsto silence tracking beacons (*google-analytics.com*,*doubleclick.net*,*youtube.com/api/stats/*,*linkedin.com/li/track*). Prevents streaming SPAs (YouTube, GCP) from stalling onnetworkidle. - Tuned Action Timings: Page load wait reduced to
0.1s, network idle wait to0.2s, action interval to0.1s. - Zero-Risk Sandbox Isolation: Completely isolated behind a single UI toggle (
Turbo Mode (Sandbox)). Unchecking it cleanly falls back to vanilla baseline behavior for side-by-side benchmarking.
- Prompt Bottleneck Elimination: Operator prompts are often underspecified or conversational. The compiler turns broad instructions into rigorous engineering blueprints in seconds.
- Automated Boolean Synthesis: Converts broad phrases like "find mid level employees of uber, microsoft, stripe and their hr and hiring managers" into exact search expressions:
"Uber" AND ("Talent Acquisition" OR "Recruiter" OR "HR Manager" OR "Hiring Manager" OR "People Operations") - Multi-Phase Architecture: Automatically formats instructions into Phase 1 (Navigation & Auth), Phase 2 (Filtering), Phase 3 (Extraction Schema Table), Phase 4 (Action & Modal Rules), and Safety Guardrails.
- Browser Use's flash mode: a 343-word system prompt instead of the full one, with no step-by-step self-check, planning or loop-breaking rules. Fine for short tasks ("open X and tell me Y").
- Off by default since 2026-10-07. On the local benchmark (
bench/), it scored 26/71 against 71/71 with Fast off. It shortened every product name in a list (0/42 over two runs), and lost fields from a list of people. It was also no faster on long tasks. Details:docs/reviews/2026-10-07-claude-accuracy.md.
- Preserves context, active tabs, and conversational thread history across multiple instructions.
- Execute chains like:
- "Search LinkedIn for AI agent engineering updates."
- "Now summarize the third result and extract author details."
- Click New Conversation → at any time to cleanly reset context and start fresh in a new tab.
- Bot Detection Bypass:
open_signin_windowlaunches a plain, non-automated Chrome profile in~/.agent-chromewithout--remote-debugging-port. Sites like Google, LinkedIn, and banking portals allow standard login without bot-blocking flags. - Cookie Flush Guarantee: The dashboard never force-terminates the sign-in window. Users log in, tick "keep me signed in", and close via the window's standard ✕ button, allowing Chrome to cleanly flush session cookies to disk without write race conditions.
- Automation Interlock: If a task starts while the sign-in window is open, the agent pauses with a handover prompt and waits for the user to close it.
- Human-Paced Scrolling:
scroll_feedscrolls incrementally, pauses up to 2.4s for lazy-loaded feeds, and senses the true bottom of the page. - Background Link Reading:
read_linkopens outbound links in an isolated CDP tab, captures the article title and text, closes the tab, and preserves the timeline scroll position without reload. - Anti-Hallucination Verification:
verify(post, page_text)compares model-reported quotes and named entities against rawdocument.body.innerTextcaptured directly over CDP. Hallucinations are either replaced with verbatim text or flagged with redunverifiedpills. - Zero-Duplicate Fingerprinting: Deduplicates posts using SHA-1 text n-grams and sanitized canonical URLs in
seen.json. - Read-Only Safety Gate:
Gate(read_only=True)blocks all likes, comments, follows, connects, or posts during digests.
- Binds to
0.0.0.0:8770by default. - Zero-Config Localhost: Workstation browser connects directly without security prompts.
- Secure Network Access: Connections across LAN / WiFi / Tailscale require the per-session token (
?token=...). Unauthorized probes receive HTTP 403 Forbidden. - Direct LAN URL with security token is printed in the terminal on startup.
- 5-Column Visual Workflow:
Queued: Dispatched tasks awaiting execution.In Progress: Active tasks navigating and executing in the browser.Needs Human: Automatically moved here during 2FA, CAPTCHA, or confirmation pauses. Returns toIn Progresswhen the operator responds.Completed: Successfully finalized tasks.Failed: Tasks that timed out or encountered unrecoverable CDP errors.
- Rich Card Properties: Automatically sets
Task(title),Status,Model,Category(Autonomous TaskvsSocial Digest),Duration (s),Steps Count,Target URL, andExecution Date. - Structured Card Content:
- Executive Summary Callout: Color-coded callout block summarizing outcomes.
- Task Instruction Quote: Verbatim prompt preserved for auditing.
- Fetched & Extracted Content: Complete body text, extracted summaries, or structured inventory.
- Execution Steps Breakdown: Expandable chronological log of all CDP actions (click, type, scroll, navigate) and visited URLs.
- 1-Click Auto-Provisioning: Enter any Notion parent page ID/URL and the agent constructs the full Kanban database schema via Notion REST API.
- Historical Back-Sync: Any past run in the Execution History can be synchronized to Notion with 1 click.
- Zero Extra Dependencies: Uses Python's native
urllib.requestagainst Notion REST APIv1(2022-06-28). Zero bloat.
Double-click Start Browser Agent.cmd or run from PowerShell:
powershell -ExecutionPolicy Bypass -File dashboard.ps1First run creates .venv and installs all dependencies automatically.
./dashboard.sh# 1. Activate virtual environment
.\.venv\Scripts\Activate.ps1 # Windows
source .venv/bin/activate # macOS/Linux
# 2. Launch dashboard
python dashboard.pyConsole URL:
- Local Access: http://127.0.0.1:8770/ (the default: only this computer can reach it)
- LAN Access (opt-in): start with
DASHBOARD_HOST=0.0.0.0, then open thehttp://<YOUR_LAN_IP>:8770/?token=<TOKEN>link printed at startup. Anyone on your network with that link can drive your logged-in browser, so use it only on a network you trust.
| Model | Mode | Vision | Optimal Use Case |
|---|---|---|---|
| MiMo v2.6 Pro (Default) | Pro | Enabled | Complex workflows, multi-step planning, high-accuracy reasoning |
| DeepSeek v4.1 Flash (Fallback) | Flash | Disabled | Ultra-fast digest extraction, text analysis, fallback execution |
| Qwen 3.8 Flash | Flash | Enabled | Spatial UI understanding, form completion |
| MiMo v2.6 Flash | Flash | Enabled | Rapid single-turn lookups |
| GLM 5.3 Flash | Flash | Enabled | General browser interaction |
The suite pauses and notifies the operator only when human agency is essential:
| Pause Trigger | Event | Action Required |
|---|---|---|
ask_human |
Ambiguity, choices, user input needed | Enter answer in dashboard input |
hand_over |
2FA, passkey, CAPTCHA, login screen | Complete action in Chrome, click Continue Execution |
confirm |
High-impact actions (payment, delete, post, submit) | Click Authorize or Decline |
Every pause triggers an audio tone, browser notification, and tab title pulse. Set NTFY_TOPIC for mobile alerts via ntfy.sh.
| Endpoint | Method | Parameters | Description |
|---|---|---|---|
/api/state |
GET | Header X-Token |
Current system telemetry, active run, queue, model configurations |
/api/run |
POST | { task, model, fallback, follow_up, fast } |
Dispatches a new autonomous browser execution |
/api/control |
POST | `{ action: "pause" | "resume" |
/api/answer |
POST | { prompt_id, answer } |
Responds to an active human-in-the-loop prompt |
/api/signin |
POST | { url } |
Launches dedicated non-automated Chrome profile for persistent login |
/api/library |
GET | Header X-Token |
Fetches watchlist targets, scheduler settings, and recent digest logs |
/api/watches |
POST | `{ action: "add" | "toggle" |
/api/new |
POST | {} |
Resets multi-turn conversational context |
/api/notion/config |
POST | { token?, database_id?, auto_sync? } |
Tests and saves Notion API integration token and target database |
/api/notion/setup |
POST | { parent_id, title? } |
Auto-provisions a new 5-column Kanban database schema under a parent page |
/api/notion/sync |
POST | { id } |
Manually synchronizes an existing execution run and steps to Notion |
For offline machines, remote servers, or home setups, use the standalone portable archive:
- Archive Path:
browser-automation-suite-portable.zip - Installation: Extract to any directory, run
Start Browser Agent.cmd. No manual setup required.
Two AI agents improve this product in parallel. Read AGENTS.md before changing anything, log every
change in CHANGELOG.md, and run python tests/test_offline.py before pushing. Design reasons,
benchmarks and the test record are in docs/ENGINEERING.md. Reviews of each other's
changes go in docs/reviews/.
This repository is maintained with clean git hygiene:
- Credentials, local Chrome profiles (
~/.agent-chrome), API keys, and runtime logs are strictly gitignored. - Remote repository synchronization:
git add . git commit -m "feat: description of update" git push origin main