agent-runtime

A Rust implementation of the Model Context Protocol (MCP) for composing AI agents, tools, and workflows. The crate provides a small runtime, event stream, and reusable step types so you can chain agents together or embed them inside other systems.

Features

• Workflow builder and runtime with step-by-step execution history
• Event stream for live progress (streamed LLM chunks, failures, completion)
• Pluggable LLM clients (OpenAI and llama.cpp HTTP endpoints included)
• Ready-made steps: agent execution, transforms, conditionals, and nested workflows
• Mermaid diagram export to visualize workflows

Quick Start

1. Install Rust 1.75+ and clone the repo.
2. Run tests: cargo test
3. Try a demo workflow (adjust the LLM endpoint in the example):
   • cargo run --bin workflow_demo
   • Other samples: hello_workflow, nested_workflow, multi_subscriber, step_types_demo, mermaid_viz, complex_viz, llm_demo, llama_demo

Architecture

• runtime: executes workflows and emits events
• workflow: builder + state for ordered steps
• agent: wraps an LLM-backed agent with prompts and tools
• step / step_impls: traits and common step implementations (agent, transform, conditional, sub-workflow)
• llm: provider-agnostic chat client trait with OpenAI and llama.cpp clients

Mermaid Output

Workflow::to_mermaid() renders a graph of your workflow (see mermaid_viz/complex_viz bins for examples).

License

Dual-licensed under MIT or Apache-2.0 at your option.