Last updated: 2026-08-17
BAML is an open-source programming language with over 9,000 GitHub stars, built by Boundary (Y Combinator W23, founded 2023) for type-safe structured LLM outputs. It compiles one .baml source file into native SDKs for Python, TypeScript, Ruby, Java, C#, Rust, and Go, and works with any model provider.
About BAML
BAML is a domain-specific programming language built by Boundary, a 10-person Seattle startup founded in 2023 by Vaibhav Gupta and Aaron Villalpando Gonzalez and backed by Y Combinator (W23). Its GitHub repository has crossed 9,000 stars and 479 forks under an Apache-2.0 license since the project's October 2023 launch. BAML exists to solve one recurring problem in LLM engineering: getting a model to reliably return data in a specific shape, and knowing at compile time, not runtime, what that shape is. Developers write .baml files that declare a function, its prompt, and its exact return type, using a TypeScript-like syntax with a Rust-style type system. BAML's Schema-Aligned Parsing (SAP) technique then coerces whatever the underlying model actually returns into that declared type, without an any escape hatch and without silent casting. The same .baml source compiles into native, typed client code for Python, TypeScript, Ruby, Java, C#, Rust, and Go, so one contract can be shared across services written in different languages. BAML also ships native Image, Audio, PDF, and Video types, so multimodal inputs do not require hand-rolled base64 plumbing, and it works against any model provider, including OpenAI, Anthropic, Gemini, AWS Bedrock, and self-hosted engines like vLLM and Ollama. The toolchain is aimed at backend and platform engineers building production LLM pipelines and agents who have outgrown ad hoc JSON parsing, and at teams whose LLM calls need to be reachable from more than one language. Tests live directly inside .baml files, a free VSCode extension gives inline prompt testing while you edit, and the public promptfiddle.com playground lets anyone try BAML in the browser with no install. Colorless concurrency (spawn and await without marking functions async) is a newer addition; Boundary's own benchmark reports a 9x wall-clock speedup on a 38GB log-scan job split across 16 parallel shards. The BAML language and local CLI toolchain are completely free: they run entirely on the developer's machine and never call Boundary's own servers. Boundary announced BAML Cloud, an optional hosted layer for observability, team controls, and governance, for later in 2026, but as of this writing pricing for that tier has not been published. There is no desktop or mobile app; BAML is installed via Homebrew, a shell script, or a PowerShell script on macOS, Linux, and Windows, and used from the command line, an editor, or the browser playground.
Pricing
The BAML language and local toolchain (compiler, CLI, VSCode extension, generated SDKs) are completely free and open source under Apache-2.0; BAML runs on your machine and does not call Boundary's servers. Boundary announced BAML Cloud, a hosted layer adding observability, team controls and governance, for later in 2026, but pricing for that tier has not been published as of August 2026.
Key Features
- Schema-Aligned Parsing: BAML's SAP technique coerces raw LLM output into the exact type declared by a .baml function, without an any type or unsafe casting, across any of the supported model providers.
- Compiles to 6 language SDKs: A single .baml source file generates native, typed client code for Python, TypeScript, Ruby, Java, C#, Rust, and Go, so one LLM contract can be called from every service in a polyglot stack.
- Native multimodal types: Built-in Image, Audio, PDF, and Video types let a BAML function accept a file directly, removing the base64 and MIME-type boilerplate most structured-output libraries leave to the developer.
- Provider-agnostic by design: BAML functions run unmodified against OpenAI, Anthropic, Gemini, AWS Bedrock, and self-hosted engines like vLLM or Ollama, so switching models does not require rewriting the calling code.
- In-editor testing: Test cases are declared directly inside .baml files and run from a free VSCode extension with an inline playground, plus the public promptfiddle.com playground for testing without installing anything.
- Colorless concurrency: spawn and await work without marking functions async; Boundary's own benchmark reports a 9x wall-clock speedup on a 38GB log-scan job split across 16 parallel shards.
Pros
- The core language and CLI toolchain are free and fully open source (Apache-2.0) with 9,000+ GitHub stars and 479 forks as of August 2026, and the toolchain runs entirely on your own machine with no calls to Boundary's servers.
- Generates native typed SDKs for 6 languages (Python, TypeScript, Ruby, Java, C#, Rust, Go) from one source file, so teams are not locked into a single language or a single model vendor.
- Built-in Image, Audio, PDF, and Video types remove multimodal input boilerplate that libraries like Instructor or raw Pydantic leave to the developer.
Cons
- Requires an extra compile step, .baml files generate client code before you can call it, which adds friction compared to zero-setup Python-native libraries like Instructor that developers can pip install and use immediately.
- BAML Cloud (hosted observability, team controls, governance) was announced for later in 2026 but remains unreleased and unpriced as of this writing, so teams that need managed monitoring today have no first-party hosted option.
- A 10-person team and an October 2023 launch mean a smaller base of production case studies, Stack Overflow answers, and third-party integrations than long-established frameworks like LangChain.
Frequently Asked Questions
How much does BAML cost in 2026?
BAML the language and its local CLI toolchain, including the compiler, VSCode extension, and generated SDKs, are completely free and open source under Apache-2.0. Boundary announced a paid BAML Cloud tier for hosted observability, team controls, and governance for later in 2026, but as of August 2026 its pricing has not been published.
Is BAML free to use?
Yes, the entire local toolchain is free with no usage caps, seat limits, or account required; it runs on your own machine and never calls Boundary's servers. The only cost you incur is whatever your chosen LLM provider (OpenAI, Anthropic, Gemini, etc.) charges for the underlying model calls.
What are the best alternatives to BAML?
Instructor is the closest zero-setup alternative for Python teams already using Pydantic, with no separate compile step. LangChain offers a much larger library of prebuilt chains, agents, and integrations if you want a batteries-included framework instead of a typed DSL. Pydantic AI is worth considering if you want Python-native structured outputs without adopting a new language.
How does BAML compare to LangChain in 2026?
BAML is a compiled, typed DSL focused specifically on getting reliable structured output from LLM calls, generating native SDKs for 6 languages from one source file. LangChain is a broader Python/JS framework with prebuilt chains, agents, and hundreds of integrations, but without BAML's compile-time type guarantees on output shape. Choose BAML for typed, multi-language contracts; choose LangChain when you want a large ecosystem of ready-made components.
How do you get started with BAML?
Install the CLI with Homebrew (`brew install baml`) on macOS or Linux, or the PowerShell script on Windows, then run `baml init` to scaffold a project and `baml ide install --code` to add the VSCode extension. From there you write a .baml file declaring a function and its return type, and BAML generates the client code for your target language. You can also try BAML without installing anything at the public promptfiddle.com playground.
Top Alternatives
- LangChain: Pick BAML if you want compile-time type safety on structured output; pick LangChain if you need its larger library of prebuilt chains and integrations.
- LlamaIndex: Pick BAML for typed, testable LLM function calls across languages; pick LlamaIndex when the job is retrieval-augmented generation over a document index.
- n8n: Pick BAML when engineers want a typed, code-first contract for LLM calls; pick n8n when the team wants a no-code visual workflow builder instead.