Introduction

What is Orca?

Orca is a declarative language for defining AI agents. Instead of writing verbose Python code with frameworks like LangGraph, you describe your agents, models, tools, and workflows in a concise, HCL-like syntax — and Orca compiles it to Python for you.

model gpt4 {
  provider    = "openai"
  model_name  = "gpt-4o"
  temperature = 0.7
}

agent writer {
  model   = gpt4
  persona = "You are a helpful writer."
}

This compiles to fully working Python code with LangGraph, complete with source mapping back to your .orca files for easy debugging.

Why Orca?

Writing AI agent systems in Python means dealing with a lot of boilerplate — setting up models, wiring tools, defining state graphs, and managing orchestration logic. As your system grows, the signal-to-noise ratio drops fast.

Orca lets you focus on what your agents do, not how to wire them up:

• Declarative — Define agents, models, and tools as simple blocks. No class hierarchies or callback functions.
• Compiled — Orca compiles to Python/LangGraph code with full source mapping, so you can debug the generated code by tracing back to your .orca source.
• Composable — Build multi-agent workflows by connecting agents with arrow syntax (agent1 -> agent2 -> agent3).
• Type-safe — The compiler catches undefined references, type mismatches, and missing fields before you ever run anything.

How it works

The Orca compiler follows a standard compilation pipeline:

.orca files → Lexer → Parser → Analyzer → IR → Code Generation

1. Lexer — Tokenizes .orca source files with line/column tracking.
2. Parser — Builds an AST using a Pratt parser with operator precedence.
3. Analyzer — Performs semantic analysis: resolves references, checks types, validates schemas.
4. IR — Converts the AST to a fully resolved intermediate representation.
5. Code Generation — Produces Python code targeting LangGraph.

The generated output includes:

• build/main.py — Your agent system as Python code.
• build/pyproject.toml — Python project config with the right dependencies.

Who is Orca for?

Orca is for programmers who want a concise, declarative alternative to writing verbose LangGraph Python code. If you find yourself copy-pasting model setup code, wiring tools to agents, and writing boilerplate orchestration logic — Orca is for you.