GizClaw/flowcraft

What it solves

FlowCraft provides a modular Go toolkit for building AI applications that are not locked into a single model provider or execution model. It solves the problem of provider lock-in and architectural rigidity by decoupling the core agent execution logic from specific inference backends and memory implementations.

How it works

The project is structured as a Go workspace with a clear layering rule: execution contracts (like engines and hosts) live in the core module, while provider adapters (drivers) and platform-specific backends stay outside the core.

• Core: Handles agent execution, graphs, tools, messages, and the runtime environment.
• Drivers: Act as adapters for providers like OpenAI, Anthropic, ByteDance, and MiniMax.
• Memory: Uses a pluggable contract system where implementations are registered by the application, allowing for flexible long-term memory strategies.
• Deployment: Applications are assembled from YAML configurations using a deployment assembly layer, which wires together agents, engines, and shared resources.

Who it’s for

It is designed for Go developers building extensible AI agents and applications that require durable execution (pause/resume), multi-provider support, and a structured way to manage AI workflows via declarative graphs or direct core package usage.

Highlights

• Durable Execution: Supports checkpoints and interrupt/resume semantics to pause and resume agents across restarts.
• Pluggable Memory: A contract-based memory architecture that allows custom implementations to be plugged in without modifying the core.
• Unified Inference: A single runtime for generation, embedding, and transcription across multiple provider families.
• Sandbox Execution: Includes built-in sandbox backends (like bwrap and seatbelt) to create execution boundaries for agents.
• Interactive Demo: Includes the forge workspace demo with a TUI and scripted tests for rapid prototyping.