%20Tokens-8b5cf6?style=flat-square&logo=sparkles&logoColor=white){kind=icon}
Autonomous Multi-Agent
Desktop Operations Platform
An enterprise-grade autonomous operating system and native PySide6 desktop suite designed for synchronized AI orchestration, 382+ senior engineering skills & agency persona synthesis, dynamic MCP execution, and live voice control.
5-Tier Enterprise Operating Architecture
Agentic Web is built on a non-blocking, decoupled 5-tier architecture ensuring fluid UI responsiveness (60 FPS), robust multi-agent background execution, military-grade security sandboxing, and resilient failover.
PySide6 Desktop & Monaco Studio Layer
Native Qt6 GUI with zero UI-thread blocking. Thread-safe Qt Signals (`token_emitted`, `status_changed`, `task_finished`) push live telemetry directly into Mission Control, Task Dispatch, Canvas IDE, and Token Manager.
Multi-Threaded QThread & QThreadPool Cluster
Dedicated worker threads managing LLM token streaming, voice processing (ASGI FastAPI + WebSocket bridge on port 8000), and Playwright Chrome DevTools Protocol (CDP on port 9222).
Security Guard & Dual-Mode Sandboxing
Deep regex command inspection intercepting hazardous OS operations (`rm -rf /`, `del /f`, fork bombs, reverse shells) with Docker container isolation and local restricted subprocess sandboxes.
Token Bucket Rate Limiting & Circuit Breaker
Granular RPM/TPM token-bucket governors, exponential backoff with full jitter for HTTP 429/5xx, and a 3-state Circuit Breaker (`CLOSED`, `OPEN`, `HALF_OPEN`) with automated model failover.
ACID SQLite WAL Database (`db.sqlite`)
High-concurrency SQLite Write-Ahead Logging (WAL) state store managing the PID process registry, task history, neural memory embeddings, and token ledgers with atomic guarantees.
Interactive UI Screen Views & Application Gallery
Multi-Agent Command Hierarchy (10 Active Models)
The platform operates on a synchronized command hierarchy where a Master Orchestrator directs atomic subtasks across specialized domain workers.
Decomposes objectives, coordinates subtasks across specialists, conducts design quality audits, and validates deliverables.
Deep technical reasoning, test-driven code implementation, complex algorithms, and rapid architectural ideation.
Final copy production, campaign packaging, client deliverables, marketing strategy, and formatted documentation.
Editorial nuance, defensive security auditing, accessibility compliance checks, and refactoring refinement.
Audience hook generation, platform engagement optimization, content formatting, and viral trend analysis.
Visual prompt composition, branding graphics, asset rendering, and aesthetic style translation.
Live web searching, source verification, comparative data auditing, and academic reference gathering.
Enterprise workflow coordination, formatting guidelines, office documentation, and spreadsheet formulas.
Technical scaling, performance profiling recommendations, API compute efficiency, and math calculations.
Fluent European multilingual translation, lightweight utility script coding, and cost-effective task delegation.
👑 Dynamic Master Leader Switching Logic
If any custom agent is registered with "is_leader": true, the platform dynamically promotes that model to LEAD_AGENT. Google Gemini is automatically placed into SPECIALIST_AGENTS (acting as a multimodal evaluator). When the custom leader is deleted or reset, Gemini automatically returns as default Master Orchestrator without data corruption.
Multi-Model Free Daily Quota Matrix (1.019B+ Tokens)
Real-time tracking of cumulative lifetime consumption ('Till Today') and daily free allowances across all leading AI providers.
| Provider & Model | Free Daily Quota / Limit | Reset Schedule | Telemetry & Auto-Routing |
|---|---|---|---|
 Google Gemini 2.0 Flash |
1,000,000,000 tokens / day | 00:00 UTC | Automatic token bucket rate limiting & ledgering |
 DeepSeek-V3 / R1 |
5,000,000 tokens / day | 00:00 UTC | SQLite WAL transaction audit trail |
 OpenRouter Free Tier |
10,000,000 tokens / day | 00:00 UTC | Model share donut visualization |
 Groq (Llama-3.3-70B) |
14,400 RPD / 500,000 tokens | 00:00 UTC | High-speed burst rate limiter |
 OpenAI GPT-4o-mini |
2,500,000 tokens / day | 00:00 UTC | Circuit-breaker failover matrix |
 Anthropic Claude 3.5 Sonnet |
100,000 tokens / day | 00:00 UTC | 7-day historical usage trend chart |
 HuggingFace Inference API |
1,000,000 tokens / day | 00:00 UTC | Token quota percentage progress bars |
 Perplexity Sonar Web Search |
500,000 tokens / day | 00:00 UTC | Live search verification token tracker |
382+ Dynamic Skills Vault & The Agency AI Personas (18 Divisions)
An extensible skills repository synthesizing 287+ The Agency AI Specialist Personas across 18 industry divisions with 24 Addy Osmani Senior Engineering Skills, automatically ingested from skills/ and dynamically injected into agent runtime prompts.
⚡ Dynamic Skills Vault Architecture & 1-Click Ingestion Engine
Agentic Web indexes skills directly from the local skills/ directory and maintains a verified high-speed registry in json/skills_registry.json (382 Total Ingested Skills).
Users can ingest any local or remote repository into the Skills Vault or Squad Roster with a single button click in the desktop app.
When an agent is assigned a task, its corresponding SKILL.md directives, vibes, and tool rules are automatically injected into the LLM system prompt.
Skills seamlessly invoke MCP tools, local terminal shells, SQLite WAL queries, and Chrome CDP viewport interactions.
⚡ MaxMasAI Harness & “Everything is a Plugin” Architecture
An autonomous evaluation, reasoning trace inspection, and hot-pluggable execution harness where every tool, model adapter, system hook, and benchmark test is a modular, hot-reloadable plugin.
Chain-of-Thought (CoT) & Multi-Turn Self-Correction
Streams step-by-step mathematical, algorithmic, and architectural reasoning traces from MaxMasAI Core and MaxMasAI Reasoner with autonomous goal decomposition and zero-shot error reflection.
Hot-Reloadable Lifecycle Hooks & 1-Click Scaffold Generator
Every feature, tool, and system capability conforms to the BasePlugin lifecycle protocol with USER_PROMPT, PRE_MODEL_CALL, and EXECUTE_TOOL event hooks. Instant hot-reloading with zero app restart.
Secure Python REPL, Shell & CDP Browser Automation Sandbox
Runs code synthesis, data transformations, and DOM interactions in an isolated execution harness with strict timeout guardrails and exit code validation.
Automated 5-Stage Evaluation & Telemetry Scoring
Continuously benchmarks algorithm generation, tool-calling precision, zero-shot RAG retrieval, and token efficiency with real-time pass/fail metrics.
🎙️ Gemini Live Voice Bridge & Native System Launcher
Real-time duplex voice streaming bridge powered by Google Gemini 2.0 Live WebSocket protocol and low-latency system application launching.
FastAPI ASGI & WebSocket Streaming (`voice/voice_daemon.py`)
Runs as a lightweight background daemon on port 8000 handling duplex audio input/output via sounddevice and numpy with seamless Gemini 2.0 Live API synchronization.
Zero-Latency Windows System Utility Dispatcher
Directly launches native desktop utilities without triggering search fallbacks: notepad, taskmgr (Task Manager), calc (Calculator), code (VS Code), explorer (File Explorer), and settings.
📐 Microsoft PowerToys FancyZones 3-Column Tiling Engine
Automated zero-overlap desktop tiling architecture powered by Win32 native APIs and Chrome CDP viewport synchronization.
Physical Screen Coordinate Resolution (`system/fancyzones_manager.py`)
Queries physical desktop workarea via SystemParametersInfoW(SPI_GETWORKAREA), dynamically extracting 3 equal zero-overlap columns with 16px margins regardless of Windows display scaling.
Win32 SetWindowPos + Playwright CDP Window Synchronization
Simultaneously snaps the PySide6 Mission Control console to Zone 1, VS Code Studio to Zone 2, and the Playwright Chrome browser to Zone 3 via Browser.setWindowBounds on port 9222.
🐙 Autonomous GitHub Issue & Crash Dispatcher
Intelligent exception interception engine connected directly to https://github.com/MaxMasAI/agentic-web/issues.
Regex Credential Redaction Barrier
Scans stack traces and exception context, scrubbing all sensitive API keys (Google API keys, OpenAI keys, DeepSeek tokens, Bearer headers, passwords) before any network transmission.
SHA-256 Crash Deduplication Engine
Computes normalized stack trace hash digests to check if an issue is already open on GitHub, automatically appending a recurrence timeline comment instead of creating duplicate issues.
⌨️ Complete Keyboard Shortcuts Cheatsheet
Master desktop navigation across Agentic Web with zero mouse movement:
| Shortcut | Destination / Action | Description |
|---|---|---|
| Alt + M | 🛰️ Mission Control | Jump to central telemetry metrics, leader status, and quick dispatcher. |
| Alt + D | 🚀 Task Dispatch | Open task execution console, latency sliders, and live terminal logs. |
| Alt + C | 🌌 Infinite Canvas IDE | Open 2D DAG graph workspace with Terminal, Agent, and Diff nodes. |
| Alt + T | 💎 Token Manager | Inspect 1.019B+ daily free tokens, 7-day trend charts, and quota countdowns. |
| Alt + S | 🤖 Agent Squads | Configure autonomous AI squads, presets, and repeated pattern auto-learner. |
| Alt + V | 💻 Monaco VS Code Studio | Full-featured Monaco code editor with split-screen HTML/CSS/JS live preview. |
| Alt + H | 🔌 MCP Service Hub | Test MCP tools live, inspect 9-field schemas, and view server registry. |
| Alt + N | 🧠 Neural Memory Vault | Search cross-mission context, rate importance, and export to RULES.md. |
| Alt + O | ⚙️ Settings & Agents | Manage custom agent schemas, PowerToys layout, and sandbox modes. |
| Ctrl + S | ⚡ Save & Auto-Reload | Saves code and triggers intelligent debounce restart via dev.py. |
Autonomous Infinite Canvas IDE (DAG Architecture)
An interactive 2D infinite workspace for designing, orchestrating, and observing autonomous multi-agent pipelines through connected DAG graph nodes.
Live Shell Execution Node
Embedded command-line node executing build tools, git operations, and test suites with live stdout streaming, ANSI colors, and exit code indicators.
Autonomous Reasoning Worker
LLM worker node displaying prompt objectives, dynamic status badges (`IDLE`, `PLANNING`, `EXECUTING`, `ERROR`), and step-by-step tool calls.
Git-Style Unified Diff Viewer
Side-by-side and unified diff viewer highlighting additions in green (`+`) and deletions in red (`-`) with 1-click file saving triggers.
Directed Bézier Data Bridges
Visual context pipelines connecting terminal stream buffers, memory context, and diff emitters seamlessly across agents.
Custom Agent Schema & API Specification
Register new custom AI models directly via Settings or by providing JSON, Python dictionary, or key-value text in the GUI.
🧪 Interactive Schema Live Validator:
Quick Start & Launch Guide
Choose your preferred environment to launch Agentic Web in under 60 seconds:
Troubleshooting & Diagnostics Matrix
| Symptom / Issue | Root Cause | Recommended Resolution |
|---|---|---|
| App window is blank / white screen | Dev server or UI thread initialization delay | Run with run.bat or python app.py. Verify PySide6 is installed in .venv. |
| Agent shows BUSY indefinitely | Network timeout or unhandled exception during browser automation | Open Inspector Drawer on the agent card in Mission Control and click Force Reset State to FREE. |
| Local endpoint connection refused | Ollama / vLLM / LM Studio service is not running | Verify local daemon is running (curl http://localhost:11434/api/tags). |
| Chrome CDP port 9222 connection retry | Remote debugging port 9222 waiting to spin up | The app automatically launches Chrome with --remote-debugging-port=9222 and retries connection automatically. |