A Python execution engine for JSON-defined AI workflows - Recipe Executor runs structured "recipes" that combine file operations, LLM interactions, and control flow into automated workflows. Perfect for AI-powered content generation, file processing, and complex automation tasks.

Recipe Executor is a pure execution engine that runs JSON "recipes" - structured workflow definitions that describe automated tasks. Think of it as a workflow engine specifically designed for AI-powered automation.

Key Features:

• 🤖 Multi-LLM Support - OpenAI, Anthropic, Azure OpenAI, Ollama
• 📁 File Operations - Read/write files with JSON/YAML parsing
• 🔄 Control Flow - Conditionals, loops, parallel execution
• 🛠️ Tool Integration - MCP (Model Context Protocol) server support
• 🎯 Context Management - Shared state across workflow steps
• ⚡ Concurrent Execution - Built-in parallelization and resource management

pip install recipe-executor

1. Create a recipe (JSON file):

{
  "name": "summarize_file",
  "steps": [
    {
      "step_type": "read_files",
      "paths": ["{{ input_file }}"]
    },
    {
      "step_type": "llm_generate", 
      "prompt": "Summarize this content:\n\n{{ file_contents[0] }}"
    },
    {
      "step_type": "write_files",
      "files": [
        {
          "path": "summary.md",
          "content": "{{ llm_output }}"
        }
      ]
    }
  ]
}

2. Execute the recipe:

recipe-executor recipe.json --context input_file=document.txt

Configure your LLM providers via environment variables:

# OpenAI
export OPENAI_API_KEY="your-api-key"

# Anthropic  
export ANTHROPIC_API_KEY="your-api-key"

# Azure OpenAI
export AZURE_OPENAI_API_KEY="your-api-key"
export AZURE_OPENAI_BASE_URL="https://your-resource.openai.azure.com/"

Recipe Executor provides 9 built-in step types:

• read_files - Read file content (supports JSON/YAML parsing, glob patterns)
• write_files - Write files to disk with automatic directory creation

• llm_generate - Generate content using various LLM providers
  • Supports structured output (JSON schemas, file specifications)
  • MCP server integration for tool access
  • Built-in web search capabilities

• conditional - Branch execution based on boolean conditions
• loop - Iterate over collections with optional concurrency
• parallel - Execute multiple steps concurrently
• execute_recipe - Execute nested recipes (composition)

• set_context - Set context variables and configuration
• mcp - Direct MCP server interactions

{
  "name": "generate_python_class",
  "steps": [
    {
      "step_type": "llm_generate",
      "prompt": "Create a Python class for {{ class_description }}",
      "response_format": {
        "type": "json_schema",
        "json_schema": {
          "name": "code_generation",
          "schema": {
            "type": "object",
            "properties": {
              "code": {"type": "string"},
              "explanation": {"type": "string"}
            }
          }
        }
      }
    },
    {
      "step_type": "write_files",
      "files": [
        {
          "path": "{{ class_name }}.py", 
          "content": "{{ llm_output.code }}"
        }
      ]
    }
  ]
}

{
  "name": "process_documents",
  "steps": [
    {
      "step_type": "read_files",
      "paths": ["docs/*.txt"],
      "use_glob": true
    },
    {
      "step_type": "loop",
      "items": "{{ file_contents }}",
      "concurrency": 3,
      "steps": [
        {
          "step_type": "llm_generate",
          "prompt": "Extract key points from: {{ item }}"
        },
        {
          "step_type": "write_files", 
          "files": [
            {
              "path": "summaries/summary_{{ loop_index }}.md",
              "content": "{{ llm_output }}"
            }
          ]
        }
      ]
    }
  ]
}

{
  "step_type": "llm_generate",
  "model": "gpt-4o",
  "provider": "openai",
  "max_tokens": 1000,
  "temperature": 0.7,
  "prompt": "Your prompt here"
}

{
  "step_type": "llm_generate",
  "mcp_servers": [
    {
      "server_name": "web_search",
      "command": "mcp-server-web-search",
      "args": []
    }
  ],
  "tools": ["web_search"],
  "prompt": "Search for information about {{ topic }}"
}

{
  "step_type": "conditional",
  "condition": "file_exists('config.json')",
  "then_steps": [...],
  "else_steps": [...]
}

recipe-executor RECIPE_FILE [OPTIONS]

Options:
  --context KEY=VALUE    Context variables (can be used multiple times)
  --config KEY=VALUE     Configuration overrides (can be used multiple times)  
  --log-dir DIR         Directory for log files (default: logs)

Examples:

# Basic execution
recipe-executor workflow.json

# With context variables
recipe-executor workflow.json --context input=data.txt output=results/

# With configuration overrides
recipe-executor workflow.json --config model=gpt-4o --config temperature=0.3

# Custom log directory
recipe-executor workflow.json --log-dir ./execution-logs

You can also use Recipe Executor programmatically:

import asyncio
from recipe_executor.executor import Executor
from recipe_executor.models import Recipe
from recipe_executor.context import Context
from recipe_executor.logger import init_logger

async def run_recipe():
    # Load recipe
    with open("recipe.json") as f:
        recipe = Recipe.model_validate_json(f.read())
    
    # Create context
    context = Context(
        artifacts={"input": "Hello World"},
        config={"model": "gpt-4o"}
    )
    
    # Execute
    logger = init_logger("logs")
    executor = Executor(logger)
    await executor.execute(recipe, context)

asyncio.run(run_recipe())

Recipe Executor provides comprehensive error handling:

• Step-level isolation - Errors in one step don't break the entire workflow
• Detailed logging - Structured logs with step-by-step execution details
• Graceful failures - Clear error messages with context information
• Resource cleanup - Automatic cleanup of temporary resources

Recipe Executor is the core execution engine of the larger Recipe Tool ecosystem:

• recipe-tool - CLI for creating and executing recipes from natural language
• recipe-executor - This package - pure execution engine
• Document Generator App - Web UI for document workflows
• MCP Servers - Integration with AI assistants like Claude

For more examples and advanced usage patterns, visit the Recipe Tool repository.

This project is licensed under the MIT License - see the LICENSE file for details.

This is an experimental project from Microsoft. For issues and examples:

• GitHub Repository
• Example Recipes
• Documentation