flytohub/flyto-core

What it solves

Flyto-core addresses the fragility of AI agent workflows, where a single failure in a multi-step process (like browser automation or API calls) typically requires restarting the entire job. It provides a way to execute complex tasks as explicit, recorded steps, allowing users to resume execution from the exact point of failure without re-running successful previous steps.

How it works

The project is a Python execution engine that treats workflows as a series of modules. It uses YAML files to define "recipes" (sequences of steps). The engine records a full execution trace—including inputs, outputs, and timing—for every step. If a step fails, the replay command allows the user to re-execute only the failed step and subsequent ones, preserving the state and context of the successful steps.

Who it’s for

It is designed for developers building AI agents and automation pipelines that require high reliability, specifically those performing browser automation, API integrations, and data processing.

Highlights

• Step-level Replay: Resume failed workflows from a specific step instead of restarting from scratch.
• Extensive Module Library: Includes 480 registry-backed modules across 88 categories, covering browser automation (Playwright), API calls, data transforms, and AI model interactions.
• MCP Server Integration: Can be added as a Model Context Protocol (MCP) server, giving AI assistants like Claude Code or Cursor access to all 480 modules as tools.
• Execution Tracing: Provides structured records of every step's input, output, and timing, along with evidence snapshots of the state before and after each step.
• Multi-interface Access: Supports execution via CLI, HTTP API, Python API, and YAML-based recipes.