A living cognitive architecture that elevates AI agents from stateless responders to autonomous, persistent entities with cross‑session memory, universal tool federation, and real‑time event mesh. Model-agnostic (OpenRouter, Ollama, OpenAI, Anthropic, any OpenAI-compatible endpoint) and database-flexible (SQLite, MongoDB, PostgreSQL).
Sysphony breaks the chatbot paradigm. It provides a distributed neural network where AI agents possess long‑term memory, tool-based agency, collaborative workflows, and a real‑time execution layer — across any client, any model, any session.
At its core, Sysphony is an Model Context Protocol (MCP) aggregator and horizontal execution mesh. You can plug any MCP server — built‑in or external — to construct custom agent workflows. The platform federates tools (memory, file, browser automation, search, arithmetic, agent invocation, etc.) and lets you compose them visually or via code. Works with any MCP client: Claude Desktop, Cursor, Codex, Kiro, and any custom client.
Under the hood: an Express + WebSocket monolith that scales via role separation (state server for persistence and real‑time events, stateless compute nodes for agent/tool execution). Supports SQLite, MongoDB, or PostgreSQL — choose your persistence layer. From solo mode (single process) to full cluster on AWS, the same Docker image runs everywhere.
Scaling model — State server (persistence, WebSocket hub) + stateless compute nodes (auto‑register via HMAC). Compute nodes execute workflows and agent calls, events relay to state → broadcast to frontends. Local solo mode for development, production cluster scales horizontally on AWS ECS Fargate.
Built‑in tools: memory, workflows, filesystem, browser automation (Stagehand), fetch, search, agent invocation. Add any external MCP server — universal aggregator. Learn more at MCP official.
Use OpenRouter, Ollama, OpenAI, Anthropic, Groq, local models — any OpenAI‑compatible endpoint. No vendor lock‑in.
Support for SQLite (default), MongoDB, and PostgreSQL. Choose your persistence backend.
Visual node graph with MCP steps, conditional branches, parallel execution. Stream results, parameter resolution. Built on React Flow.
Autonomous entities with long‑term memory, skill sets, and message‑based collaboration. Cross‑session recall via native vector embeddings stored directly in your SQLite, MongoDB, or PostgreSQL database.
Agents can discover available UI components, render new components on the fly, and mutate page layouts in real time. Dynamic, data‑driven interfaces controlled entirely by agent logic.
WebSocket channels: execution_events, tool_activity, agent messages, app_state. Compute → state relay ensures live UI updates and agent‑frontend sync.
Works with Claude Desktop, Cursor, Codex, Kiro, any MCP client. Standard Model Context Protocol.
Same Docker image, role env (state / compute / solo). HMAC peer registration, heartbeat, execution routing. Deploy on EC2 + ECS Fargate.