FlowCraft
A modular Go toolkit for building AI agent applications with long-term memory, knowledge retrieval and voice, usable as a library, a background daemon, or a real-time pipeline, without tying the code to one model provider.
FlowCraft's core module defines agent execution, graphs, tools, models, messages, inference, memory contracts, an event bus, telemetry, workspaces, sandboxing and delegation as a single platform layer; graphs are one optional way to structure an application rather than a required architecture, and code can also use the core packages directly. Provider driver modules adapt OpenAI, Anthropic, ByteDance and MiniMax to this common interface, so a deployment pointed at any endpoint speaking one of those wire protocols, such as DeepSeek or GLM riding the OpenAI driver, is named by the endpoint's provider id rather than the module that implements it.
Its memory contracts are implemented by pluggable, independently registered memory backends, mirroring how each inference provider is a registered factory around the generic inference package. SQLite-backed checkpoints and sandbox backends live under their own module, and a runnable local workspace demo with a terminal UI ships as an example. Each library layer is an independently versioned Go module, so an application only needs to adopt the layers it uses.