Boundary ML

BAML is an open-source, Turing-complete programming language designed for agentic coding—software written by agents and humans together. It feels familiar to TypeScript developers, but where TypeScript optimizes for human productivity, BAML optimizes for agent safety. It keeps types meaningful at runtime, has no `any` or unchecked casts, and makes invalid states hard to represent. You write prompts as .baml files, and BAML compiles typed functions that call LLMs, guaranteeing structured outputs and catching schema mismatches at compile time. It generates native client code for Python, TypeScript, Go, Java, C#, Rust, C++, Swift, and the web. The toolchain is designed for real agent workflows. The CLI includes `baml describe` for AST-aware code discovery, `baml run` to execute any function, and `baml test` for CI/CD integration. The local playground lets you iterate on prompts and validate outputs. BAML supports green threads with `spawn` and `await` for concurrent execution, typed errors with exhaustive pattern matching, and native LLM functions, tests, testsets, and evaluations. Built-in stdlib packages for OpenAI and Anthropic provide retry, fallback, round-robin, and streaming clients. A recent nightly replaced the sys_llm shim with native provider clients for OpenAI, Anthropic, Google, and others, and renamed the OpenAI client to openai.ResponsesClient. BAML integrates with major LLM providers: OpenAI, Anthropic, Google AI, AWS Bedrock, Azure OpenAI, Llama, Mistral, Ollama, Groq, and Hugging Face. It runs on any cloud or entirely on your machine, and the local toolchain is free and open source. BAML Cloud—with observability, team controls, and governance—is planned for later this year. Cloud pricing isn't published yet. BAML is pre-1.0 but already production-ready for data extraction and type-safe agent workflows. If you're a developer tired of unpredictable JSON from raw LLM prompts, BAML gives you compile-time guarantees and a testing harness that catches regressions before deployment.