I built OpenMolt, an open-source framework for creating programmatic AI agents in Node.js.
Instead of wiring complex workflows (that can break), you define agents in code with instructions, tools, integrations, and memory. The agent then plans steps and executes actions like calling APIs, updating data, or interacting with services. When things break or change, the agent can adapt instead of failing like static workflows.
The goal is to make agents behave more like real software systems than prompt scripts.
It includes: - a planning/execution loop - tool and API integrations - memory support - scheduling - a CLI runner
One thing you're probably worried about: Security With OpenMolt, agents are capability-scoped: they can only access the tools, integrations, and permissions explicitly granted to them, which helps keep automation predictable and secure. Also, the LLM only sees tool names, never raw credentials.
Would love feedback on the architecture and developer experience.