Problem

Building reliable applications around large language models is still too much like stitching together raw prompts, application code, parsers, retries, and testing infrastructure. Boundary’s documentation describes prompts as lacking type safety, previews, and linting, while structured-output workflows often force developers to execute a Python/Pydantic environment or build a separate microservice just to test a small prompt change. The pain is primarily engineering time and operational risk: teams must manage complex model and retrieval choices, hallucinations, malformed outputs, and confidently incorrect answers that can damage customer trust and create compliance or support costs.

The killer use case is turning unreliable LLM calls into production-grade application functions, especially for AI agents, RAG chatbots, document and resume extraction, classification, and tool calling. These applications need predictable schemas and error handling rather than prose responses, and BAML is aimed at the point where a large share of business logic is being delegated to prompts or agents.

Product / Service

Boundary’s core product is BAML, an open-source, domain-specific programming language for agents and structured LLM outputs. Developers define typed functions, schemas, prompts, and model clients in BAML; the toolchain generates client code for languages such as Python, TypeScript, Go, and Ruby. Generated code calls the selected LLM, injects the output schema into the prompt, parses the response, repairs broken JSON, applies retry logic, and returns a typed object that can be validated and consumed by the host application. BAML also provides a local playground, prompt previews, tests, evaluations, concurrency, and tools for observing LLM inputs and outputs.

The delivery model is currently local-first: the language and toolchain are free and open source, run on the developer’s machine, and can be adopted incrementally alongside an existing framework without sending requests through Boundary’s servers. Boundary says a hosted BAML Cloud is planned to add observability, team controls, and governance, creating a potential commercial layer while preserving the free local runtime. The benefit is less boilerplate and faster iteration, with stronger guarantees around structured outputs and a common interface that lets teams use different models and programming languages.

Market

Boundary competes in AI developer tools and the emerging programming-language and infrastructure layer for LLM applications and agents. Its closest substitutes are not only other structured-output libraries but also agent and orchestration frameworks such as LangGraph, CrewAI, AutoGen, DSPy, and the OpenAI Agents SDK, plus type-oriented tools such as PydanticAI and Instructor. BAML’s differentiation is to move beyond a library inside one host language: it offers a dedicated language, generated SDKs across multiple languages, native testing and evaluation, and tooling designed for agents to inspect and modify code. It can also coexist with existing LLM frameworks, so the competitive relationship is partly complementary.

Boundary appears to be an early but meaningfully adopted, seed-backed company rather than a purely pre-revenue experiment. It was founded in 2023, participated in Y Combinator’s Winter 2023 batch, and PitchBook reports a $1.5 million seed round and $2 million raised in total. The company’s 2025 account described multiple Fortune 500 companies, government agencies, and startups using BAML, more than 2,400 Discord members, and growth from 1,000 to 7,000 GitHub stars; the repository evidence now shows 8,744 stars and 463 forks. Boundary has also reported paying customers historically, although current revenue is not publicly quantified and the core local product remains free while its planned cloud offering has not yet published pricing.