What is hermes-synapse?

Hermes is a visual framework for building networks of AI agents that plan their own steps and coordinate through a strict directed graph, structured so a task branches into sub tasks but never loops back on itself, avoiding the runaway back and forth that some other agent frameworks fall into. Users wire agents together on a drag and drop canvas rather than writing code by hand, and a root orchestrator called Jarvis sits at the entry point, deciding how to route each incoming request down through sub orchestrators and sub agents. Each agent can be given a set of skills, which are pluggable tools such as web search, a Python code sandbox, weather lookups, a calendar integration, or a local memory system backed by the Qdrant vector database. The project ships with a set of ready made agents out of the box: a search agent, a code engineer that can self correct its own scripts, a data analyst that builds charts, a scheduler, a market monitor for stocks and crypto, a daily planner that connects to Google Calendar and Todoist, a system operations agent, and a football analyst. Agents missing a required API key still start, they just report plainly what needs to be configured. Security is enforced through what the project calls Permission Intersection: a child agent can only use the tools that both it and its parent were given, so a sub agent cannot gain access to something like shell execution just because it happens to be configured with that skill, if its parent lacks it too. This is meant to stop one agent from quietly escalating its own privileges as tasks move down the chain. The backend is a FastAPI Python server with a React and Vite dashboard, and it can connect to any OpenAI compatible language model provider, including OpenRouter's hosted models or a locally run model through Ollama. Setup is done through Docker Compose, using a single environment file for configuration, and the project supports SQLite by default or PostgreSQL for production use. The code is released under the MIT license.