This Rust-based engine provides a YAML-native declarative system for defining, orchestrating, and executing complex agent workflows. It targets developers building multi-agent systems or those seeking a robust backend for visual agent IDEs like RicePrompt, offering a streamlined approach to managing LLM calls, data processing, tool integration, and multi-agent coordination through a single configuration file.

How It Works

The engine parses a single YAML file that describes the entire workflow, including nodes, edges, prompts, data sources, and tool integrations. It resolves dependencies and executes the defined graph, supporting various node types such as LLM generation, Rhai scripting for transformations, iterators, multi-agent supervisors, subgraphs, and data connectors. This declarative, YAML-first approach simplifies the definition and management of intricate agent behaviors.

Quick Start & Requirements

To run a workflow, you define it in a YAML file (e.g., hello.yaml) and execute a Rust program that uses the riceprompt-engine crate. The example requires an OPENAI_API_KEY environment variable. More runnable examples are available in the examples/ directory. The authoritative specification for the YAML workflow format can be found at docs/FLOW_SPEC.md.

Highlighted Details

• YAML-native: Entire workflow specification resides in a single declarative file.
• Multi-provider LLM Support: Integrates with OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Zhipu, Moonshot, MiniMax, xAI, Huoshan, and any OpenAI-compatible endpoint, with first-class support for streaming, tool calling, and structured output.
• Rich Node Types: Includes generate, transform (Rhai scripting), iterator, supervisor (multi-agent routing), subgraph, data_connector, skill_set, and mcp/mcp_tools.
• Built-in Data Connectors: Supports PostgreSQL, MySQL, MongoDB, Redis, Qdrant, S3-compatible object storage, and REST APIs.
• Harness Layer: Allows workflow-level instructions with persistent memory support.
• Self-Describing Results: Execution results can include the source YAML for easy visualization of topology and per-node outcomes.
• Checkpoint/Resume: Enables pausing and resuming long-running workflows.

Maintenance & Community

Contributions via issues and PRs are welcome, with requirements for code formatting (cargo fmt) and linting (cargo clippy). Changes affecting the YAML surface necessitate updates to docs/FLOW_SPEC.md. The project is closely related to RicePrompt, a visual agent IDE built upon this engine. Specific community channels like Discord or Slack are not listed.

Licensing & Compatibility

The project is dual-licensed under either the Apache License, Version 2.0 or the MIT License, offering flexibility for integration. These standard open-source licenses generally permit commercial use and linking within closed-source projects, provided the terms of the chosen license are met.

Limitations & Caveats

The engine is currently in a 0.1.x release series, indicating that its API may be subject to change between minor versions as the workflow specification stabilizes. Users requiring stable behavior should pin an exact version of the riceprompt-engine crate in their dependencies.