GizClaw/flowcraft

Production-grade Go SDK for building AI agents with long-term memory, knowledge retrieval, and voice — runnable as a library, a daemon, or a real-time pipeline.

What it solves

FlowCraft is a modular Go toolkit designed to build extensible AI applications without locking the developer into a single model provider or execution model. It provides a standardized way to handle agent execution, long-term memory, and plugin-based application assembly, allowing developers to swap providers and backends easily.

How it works

The project is structured into several layers:

  • Core: Defines the execution contracts for agents, graphs, tools, and memory. It includes a declarative DAG engine for compiling graphs into agent engines.
  • Craft: An assembly layer that uses craft.yaml definitions and compile-time capabilities to build applications. It manages a plugin host where plugins act as MCP (Model Context Protocol) servers.
  • Drivers & Backends: Provider-specific adapters (e.g., OpenAI, Anthropic) and platform implementations for memory (SQLite) and secret storage.
  • Runtime: Manages leased, interruptible streaming sessions and handles the deployment of resources defined in YAML documents.

Who it’s for

Go developers building complex AI agents or applications that require provider-agnosticism, durable execution (pause/resume), and a plugin-based architecture for extending functionality.

Highlights

  • Pluggable Memory: Memory is treated as a pluggable implementation registered behind core contracts.
  • Durable Execution: Supports checkpoints to pause and resume agents across restarts and uses a subject-routed event bus for structured telemetry.
  • Unified Inference: A single runtime for generation, embedding, and transcription across multiple providers (OpenAI, Anthropic, ByteDance, MiniMax).
  • MCP Integration: Plugins operate as MCP servers, and the host exposes its own primitives as MCP tools for plugins to call back into.
  • Sandbox Execution: Includes built-in sandbox backends (bwrap, seatbelt) to isolate agent execution boundaries.

Related

  • Project
    bytedance/flowgram.ai

    FlowGram is an open‑source visual workflow framework (bytedance/flowgram.ai) for building AI‑agent pipelines. It offers a web SDK and a native desktop app with drag‑and‑drop canvases, built‑in LLM, HTTP, and code nodes, scoped variables, form generation, debugging, and versioned publishing. Quick start via `npx @flowgram.ai/create-app@latest` and extensive docs at flowgram.ai.

  • Project
    fastclaw-ai/fastclaw

    FastClaw is a Go‑based, single‑binary runtime that lets you create, manage, and run multiple AI agents with their own personalities, memory, and skills. It supports any OpenAI‑compatible LLM, sandboxed tool execution, channel bots (Telegram, Discord, Slack, etc.), cron‑style scheduling, and a full web UI + CLI + OpenAI‑compatible API. Data is stored in SQLite (or Postgres) and configuration is env‑only. The project is source‑available under a community license.

  • Project
    withastro/flue

    A programmable TypeScript harness for building autonomous agents with built-in support for secure sandboxes, durable sessions, and tool integration.

  • Project
    DotHarness/dotcraft

    DotCraft is an open‑source AI‑agent platform you run locally. It provides a desktop app, CLI, Android client, and SDKs (TypeScript/.NET) so the same agent can plan, execute, and coordinate sub‑agents across devices, integrate with IDEs, chat platforms, and embed into your own applications.

  • Project
    nextlevelbuilder/goclaw

    A multi-tenant AI agent platform built in Go that provides a unified gateway for 20+ LLM providers and 7 messaging channels with built-in multi-agent orchestration.