Open-source TypeScript framework for rapid application development. A production-ready foundation with zero vendor lock-in and full control over your own code.

Stack: Vue 3 and Vite on the front, Express and Prisma on the back, Redis for queues and caching, Docker for one-command deployment.

Out of the box: authentication (JWT, OAuth, sessions, 2FA), role-based access control (CASL), background jobs (Redis queues, cron, events), a CMS module (content, pages, HTML cache), communications (email, WebSocket), and 50+ Vue UI components.

Plugin-based architecture: application code stays isolated from the core, so core updates do not break customisations. Each feature lives in its own plugin module and can be enabled or disabled by config.

Built for pairing with AI coding assistants: structured prompts generated from models and routes, strict guardrails through generated types and validators, context capsules for Claude Code, Cursor and ChatGPT, and a review trail with lint and tests over generated code.

Running in production across multiple projects. Self-hosted only: deploy to any VPS, cloud, or Kubernetes.

Three minute walkthrough:

Site: https://typus.dev

An AI server a business installs on its own Linux box. It sits between day to day operations (email, messaging, documents, tasks) and whichever model providers the business chooses, so orchestration logic, agent definitions, audit trail and customer data stay on their infrastructure rather than with a SaaS vendor.

Architecture: an AI Gateway as the single egress point for every external LLM call, so providers (OpenAI, Anthropic, Ollama) are swappable without touching application code. Semantic memory on PostgreSQL with pgvector. YAML-defined workflow orchestration with crash recovery. Task tracking with approval gates and a full event log. Email and messenger channels for human handoff.

Open source components: Mesh (semantic memory), Rein (workflow orchestrator), Screenbox (isolated agent desktops with a real Chromium browser).

Running against real business operations. One inbound phone call, handled end to end, moving work through the whole node:

Site: https://smart-node.app

Turns a topic, a script, or a document into a finished hand-drawn explainer video with voiceover. Real drawing, not slideshow fades: strokes are traced and animated on scene by scene, the way a person sketches at a board.

Fully headless pipeline. Text goes in, the system writes the narrative, splits it into scenes, generates the artwork, traces it to SVG strokes, animates the draw-on, synthesises the voiceover, and assembles the final MP4 with FFmpeg. No manual editing step anywhere in the chain.

Stack: TypeScript, Vue 3, Express, FFmpeg, Docker. Self-hosted, built on the Typus framework.

Media handling throughout: image generation, SVG rasterisation and stroke tracing, frame assembly, audio and video muxing, encoding.

Site: https://explain.ink

Self-hosted semantic memory system that gives AI agents persistent, searchable memory with automatic tagging. Agents store and retrieve documents by meaning, not keywords.

Built on PostgreSQL with pgvector for embedding-based search. Documents get date and source tags automatically on save. Type, project, and topic tags are inferred from similar existing documents after indexing.

Multi-workspace support for separating different projects or teams. Full document version history. CLI tool and MCP server for direct integration with Claude Code or any MCP-compatible client.

Used in production as a shared knowledge layer between 5+ AI agents working on the same projects. Agents write findings, decisions, and worklogs. Other agents search and build on that context across sessions.

Works fully offline with local ONNX embeddings. No external API calls needed for search.

GitHub: https://www.twine.net/signin

Self-hosted knowledge base platform where teams upload documents and get answers through a chat interface. Users ask questions in natural language and get responses grounded in their actual documentation, not generic AI answers.

Supports PDF, Markdown, and text documents. Semantic search finds relevant content across the entire knowledge base. Chat interface provides sourced answers with references to specific documents.

Self-hosted on Docker, all data stays on the client’s infrastructure. No external API dependencies for document storage or search. Works with Claude and other LLM providers for the chat layer.

Built for teams that need internal documentation accessible through conversation without sending sensitive content to third-party SaaS platforms.

Site: https://www.twine.net/signin

Self-hosted platform that gives each AI agent its own isolated Linux desktop with a real Chromium browser. Agents see the screen, click, type, navigate tabs. You watch them work in real time via built-in dashboard and take control when needed.

Docker-based, MCP-native. Works with Claude Code, Claude Desktop, Cursor, or any MCP-compatible client. Each desktop is a container with full isolation, no bind mounts or host access.

Features: live screenshot streaming, Chrome DevTools Protocol semantics for faster element interaction, snapshot and restore of desktop state, knowledge compilation from past sessions, human-in-the-loop control.

Around 2GB RAM per desktop, no GPU required. Runs on any Linux server or Windows with WSL2.

Site: https://screenbox.dev
GitHub: https://github.com/dklymentiev/screenbox
Demo:

Open-source workflow orchestrator for AI agents. Define multi-step processes in YAML, assign each step to a specialist agent, and run with automatic crash recovery via SQLite.

Each block in a workflow can be an LLM call (Claude, GPT, Ollama, OpenRouter) or any executable script. The orchestrator handles dependencies, data flow between blocks, and error handling. Parallel execution, conditional branching, revision loops, and tag-based routing supported.

Three execution modes: sync (CLI, wait for result), async (daemon, poll for progress), step mode (run N blocks per invocation, save state, exit).

Built-in MCP server lets you trigger workflows directly from Claude Desktop, Cursor, or Claude Code. Optional daemon mode with WebSocket for live updates.

MIT licensed. Five working examples in the repo.

GitHub: https://github.com/dklymentiev/rein-orchestrator
Demo: