GizClaw/flowcraft

What it solves

FlowCraft is a production-grade toolkit for building AI agents in Go. It addresses the complexity of moving from a simple LLM prompt to a production-ready system by providing a layered architecture that separates the agent logic from the runtime engine, memory management, and provider integrations.

How it works

FlowCraft is organized into several modular layers:

• SDK: Provides core primitives like agents, DAG executors, and tool management.
• Memory: Implements long-term recall using a hybrid retrieval system (BM25 + vector + entity) and manages conversation history.
• SDKX: Offers drop-in providers for various LLMs (OpenAI, Anthropic, DeepSeek, etc.) and embedding models.
• Vessel & Vesseld: An in-process runtime (vessel) and a standalone daemon (vesseld) that allows agents to be deployed via declarative YAML, providing lifecycle management and an HTTP/SSE control plane.
• Voice: A real-time pipeline for STT (Speech-to-Text) $\rightarrow$ LLM $\rightarrow$ TTS (Text-to-Speech) with built-in VAD (Voice Activity Detection) and WebRTC support.

Who it’s for

It is designed for Go developers who want to build scalable, production-grade AI agents with long-term memory and voice capabilities without having to manually wire together disparate libraries or rely on Python-based frameworks.

Highlights

• Hybrid Memory: Uses Reciprocal Rank Fusion to combine BM25, vector, and entity retrieval for more accurate recall.
• Declarative Deployment: Deploy agents as a standalone daemon using YAML configurations.
• Provider Portability: Switch between different LLM providers by changing a single YAML field.
• Durable Execution: Supports pausing and resuming agents via a Checkpoint store.
• Integrated Voice Pipeline: Built-in support for barge-in, endpointing, and WebRTC for real-time voice interactions.