# BAML

> A domain-specific language for type-safe LLM functions, with generated clients and schema-aligned parsing.

BAML is a small domain-specific language for defining LLM functions with typed inputs and outputs. You write the function and schema once in .baml files and generate type-safe clients for Python, TypeScript, and more; its schema-aligned parser reliably coerces messy model output into your types.

Website: https://www.boundaryml.com

BAML (by BoundaryML) takes a different angle on structured LLM output: instead of a library in one language, it's a small **domain-specific language**. You define an LLM function — its prompt, typed inputs, and typed output — in a `.baml` file, and BAML generates **type-safe clients** for Python, TypeScript, and other languages. Its **schema-aligned parsing** is built to coerce the imperfect, almost-JSON that models often emit into your declared types.

It is aimed at teams who want their LLM calls to be first-class, version-controlled, type-checked artifacts shared across a polyglot codebase, rather than prompt strings scattered through application code. The DSL also gives you a testing playground for prompts.

## Highlights

- **Typed LLM functions** — declare inputs, outputs, and the prompt in one place; get compile-time types in your app.
- **Multi-language clients** — generate idiomatic clients (Python, TypeScript, and more) from the same definition.
- **Schema-aligned parsing** — robustly parses real-world model output, including partial/streaming and malformed-ish JSON.
- **Prompt as code** — `.baml` files live in version control with tests and a playground.
- **Provider-agnostic** — works across model providers.

## In an AI-assisted workflow

```baml
class User { name string  age int }

function ExtractUser(text: string) -> User {
  client Claude
  prompt #"Extract the user from: {{ text }} {{ ctx.output_format }}"#
}
```

Generate clients, then call `ExtractUser(...)` from Python or TypeScript and get a typed `User` back.

> [!TIP]
> BAML shines when the same prompts are consumed by multiple services/languages, or when you want prompts under test and review like any other code. For a single-language, in-code approach, [Instructor](/tools/instructor) is lighter weight.

## Good to know

BAML is open source and free; you pay your model provider for tokens. It introduces a build step (generating clients from `.baml` files), which is the trade for cross-language type safety. See [Structured Output vs JSON Mode vs Function Calling](/guides/concepts/structured-output-2026) for when typed output is worth it.

---

_Source: https://agentscamp.com/tools/baml — Tool on AgentsCamp._