AvDev

Underneath the builder runs a DAG-based task graph, not a single AI call. A Controller manages the lifecycle. A Planner decomposes the request into sub-tasks with dependencies. Six worker roles execute — Research, Code, Browser, Test, Debugger, Deployment — each spawned for what it's good at. An Agent Registry tracks status, iteration count, token and credit budget, and output, with hard limits that stop runaway agents. Verification is a first-class layer: an Observation pipeline captures screenshots, console logs, network activity, and DOM state after every action; a Validator checks structural correctness; a Critic red-teams another agent's work for bugs, security flaws, and design drift; a Gate blocks the next DAG node until verification passes.

AvDev keeps context across long sessions through four tiers of memory. Working memory holds the current task's active state. Global memory carries cross-project knowledge, preferences, and learned patterns. An immutable Execution Log records every action, output, error, and rollback point. Hierarchical RAG memory retrieves against embeddings for design preferences, past decisions, and reusable code. Safety is enforced with iteration, spend, and concurrency caps, a Rollback Manager that reverts to the last known good state on failure and retries with a different strategy, and a Silent Fix Agent that watches build and runtime errors in the background and patches them without user intervention.

A PlatformCapabilities registry tracks which services are live. AvDev Cloud is a zero-config backend — Postgres with RLS, email and OAuth auth with JWT sessions and XOR-encrypted profiles, Deno edge functions, private and public storage buckets, realtime, and secrets. AI runtime introspection lets the model read table schemas, invoke edge functions, list secret names, and browse storage under a scoped JWT. AvDev AI covers chat, image, video, and embeddings with model routing and an offline mode for private codegen.

AvDev goes where most codegen tools stop. Web3 — wallet connection with JWT-backed sessions, smart contract deployment and verification, DeFi with swaps and staking, NFT minting with IPFS pinning, and a transaction simulation pre-flight with exact token approvals and no fake bytecode fallbacks. AGI OS — a desktop agent that drives the local machine through an app launcher, sandboxed bash executor, local browser, file executor, and screen executor, using $done and CONFIRMED signals for safe handoff. Video Production — a Remotion and HuggingFace pipeline that generates marketing videos, explainers, and ads with structured scene composition. Ecosystem OS — multi-app workspaces where a family of apps shares data, auth, and workflows and can message across itself.

A large internal toolset the agents share and users can reach through chat: web search, Puppeteer browser automation, unified codebase search across code, chat history, and error logs, security and performance analyzers, deep research, image generation, image-to-video and audio synthesis, document parsing and generation, cloud read and write with edge function invoke and secret listing, blockchain price feeds and contract helpers, GitHub repo lifecycle with webhook ingestion, Shopify storefront hooks, and an MCP connector for external knowledge sources.

The system watches how you edit and learns style preferences — colors, spacing, typography — to apply to future generations. A Design Token Generator produces semantic palettes, spacing scales, and typography. A background Style Learning Worker suggests improvements as patterns emerge. A Visual Canvas Engine bridges vibe-coding: edit layouts visually and the AI transpiles changes into code. Select-to-Chat lets you click any UI element to open a chat scoped to that specific component. Response templates keep AI output structurally consistent, and auto-sync hooks keep hooks and state coherent across the codebase.

Security runs as its own layer — an RLS scanner that validates every table has row-level security, a dependency scanner that checks npm packages for known CVEs, a code scanner that catches secrets leakage, SQL injection, and XSS, and a report generator that produces human-readable audit output. XOR-encrypted profiles keep sensitive data at rest. OAuth uses hardened canonical redirects. On the commercial side, every AI operation costs metered credits with per-request and per-token tracking, throttling to prevent abuse, tiered subscriptions with monthly allotments, built-in referrals, and Paddle and Stripe for global billing. One-click publish ships to a live URL with custom domain and full DNS, sitemaps and JSON-LD, SEMrush-backed keyword and backlink research, and Search Console verification.

In private beta with a small group of people I trust to tell me when it's bad. Which is often.