dottxt

dottxt (branded ".txt") develops structured-generation, or constrained-decoding, software that makes large language model outputs conform to a specified schema at generation time rather than through post-hoc parsing, retries or validation loops. Its technology enforces JSON Schema, regular expressions and context-free grammars, and is also applied to function calling. The company frames a schema as a contract that its software enforces, with the stated goal of making AI systems compose like conventional software across layers from token generation to agents.

The company is the maintainer of Outlines, an open-source Python library for structured generation that works across model providers and serving stacks including OpenAI, Anthropic, Gemini, Ollama, vLLM, SGLang, TGI, Transformers, llama.cpp and MLX-LM. Commercially, dottxt offers three product lines: the api.dottxt.ai platform, an OpenAI-compatible, pay-per-token hosted API providing constrained decoding over open-source models; drop-in replacements for self-hosted inference servers (dotvllm, dotsglang, dottensorrt_llm) targeting teams running vLLM, SGLang or TensorRT-LLM; and composable libraries for inference providers — dotjson for JSON Schema, dotgrammar for arbitrary context-free grammars and dotlambda for function calling. It also offers a schema review/audit service that reports what breaks under generation and compliance rates before and after applying constraints.

According to press coverage, dottxt positions its commercial offering as model-agnostic with more features and better performance than the open-source Outlines project it grew out of, and its founders have emphasized that constraining generation does not have to degrade model performance.

Founding story

Co-founders Rémi Louf, Dan Gerlanc and Brandon Willard had worked together for roughly four years, most recently at New York-based AI firm Normal Computing, where they hit problems getting GPT-4 to extract data in a systematic order without extensive manual work. Drawing on backgrounds in statistical modelling, Bayesian statistics and compiler technology, they built a structured-generation solution and open-sourced it as Outlines, which was originally developed at Normal Computing by Rémi Louf and Brandon T. Willard before being spun into the company.

Business model

Commercial licensing and hosted services layered on an open-source core: dottxt maintains the Apache-2.0 licensed Outlines and outlines-core projects while selling commercially licensed structured-generation libraries, self-hosted inference-server replacements, and a hosted API. Enterprises and developers are the intended paying customers, with schema review offered as an entry point.

Pay-per-token usage of the api.dottxt.ai platform, plus commercial licenses/enterprise support for structured-generation libraries and drop-in inference server components sold to self-hosting teams and inference providers.

Traction

Outlines has been cited at over 2.5-3 million downloads by October 2024 (with roughly 600,000 in the preceding month) and "65M+ downloads" on the company's site; the GitHub repository shows about 15,700 stars and 868 forks, and outlines-core about 309 stars. Adoption is reported among hundreds of organizations and by NVIDIA, Cohere, Hugging Face and the main LLM serving frameworks. Headcount grew from eight or nine people in mid-2024 to 17 by the end of October 2024.

Latest developments

The company's current product line spans the pay-per-token api.dottxt.ai platform (OpenAI-compatible endpoints, described as in early access), drop-in replacements for vLLM, SGLang and TensorRT-LLM inference servers, and the dotjson, dotgrammar and dotlambda libraries for inference providers. Messaging has broadened from per-call structured outputs to reliability for agent infrastructure, and the company promotes a schema-audit service reporting compliance rates before and after constraints. Public repositories including Outlines, outlines-core and the official dottxt Python client remained under active development into 2026.