The AI-native coding environment that doesn't just assist — it thinks, learns, debates, and improves alongside you. Powered by a custom neural architecture. Runs entirely on your machine. No cloud. No subscriptions. No limits.
Portal IDE is not a chatbot wrapper. It runs a custom recurrent neural architecture with O(1) memory complexity — processing infinite context at constant memory cost. Every intelligence metric is earned through real training, not seeded with placeholder data.
A custom Minimal Gated Recurrent Unit architecture with O(1) memory complexity. No transformer attention caps. Processes infinite context at constant memory cost. Trains and runs fully locally with an optional log-space Triton kernel for GPU acceleration.
Two AI agents — TinyCPUDevAI and MentorAI — argue over your problem in a structured deliberation loop. Results are synthesised into a joint answer that dramatically reduces hallucination on complex reasoning tasks. The most sophisticated AI debate system in any IDE.
Multi-core cognition splits complex tasks across parallel MinGRU agents. Each "head" specialises in logic, syntax, creativity, or code review. Results are merged with confidence-weighted voting via the CognitiveCore engine.
Every interaction is scored, categorised by intent, and stored with language context. The Shadow Orchestrator fine-tunes on wins and losses automatically during idle periods — the IDE genuinely gets smarter the more you use it.
A shared publish/subscribe memory layer that persists tagged facts across all agents in real time. Any agent can post, query, or clear the blackboard as a first-class tool. A dedicated live feed dock shows all entries with timestamps and agent source tags.
A full Reason + Act agent loop that plans, observes results, reflects, and iterates until a final answer is reached. Every reasoning step is displayed in a dedicated Reasoning Panel dock — full transparency into the AI's decision-making process.
Portal IDE ships as a fully self-contained development environment — tabbed editor, embedded terminal, file explorer, Git panel, debugger, and more. Everything you need, nothing you don't.
Portal IDE ships with an 8-source training pipeline that feeds a self-improving AI brain. Train from web documentation, videos, books, GitHub repositories, and more — all from inside the IDE.
Headless Playwright + BeautifulSoup fallback scraping 120+ curated sources across 15 categories — language docs, frameworks, DevOps, databases, ML/AI, and more. Robots.txt compliant with adaptive rate limiting.
Local Whisper-powered video ingestion extracts spoken transcripts and OCR code-on-screen from MP4/MKV/AVI files. No API cost. Transcripts feed directly into the training corpus.
Smart PDF, EPUB, Markdown, and plain-text parser that chunks content by heading structure and preserves code-block continuity. Metadata stored alongside content for RAG retrieval.
Clone any public Git repository or scan a local path — extracts function signatures, docstrings, dependency trees, and build context. Language auto-detected per file across 30+ languages.
Structured training progression: Token Prediction → Code Completion → Code Generation → Code Review. Advances automatically when mean loss stays below threshold. Session stats written to the database at every epoch.
9 interactive matplotlib charts auto-refreshing every 4 seconds — intelligence gauge, 3D training surface, media mix pie, win/loss stacked bar, and thinking speed histogram. Every metric earned through real training.
Portal IDE ships with a security and resilience layer that most commercial products don't have. Every subsystem is hardened, every failure mode is handled, and every file write is atomic.
We're putting the finishing touches on the release build. Be the first to know when Portal IDE launches.
✉ Get Notified at LaunchAlso available: TinyCPUDevAI · Monster Maneuvers