B4run

B4.run is an open source TypeScript framework for building AI agents in which files are features. A single command scaffolds a new project, and each agent is simply a folder: an index file picks the model and writes the system prompt, and every TypeScript file placed in its tools folder becomes one of the agent's tools. Adding further files gives the agent more abilities, such as a plan file for planning, a memory file, skills written as markdown, subagents in their own folders, and evaluation files that check behaviour. It is aimed at TypeScript developers who want agent code they can read and review rather than configuration hidden inside a framework. Tool schemas come straight from the code: the file name becomes the tool name, the JSDoc comment becomes its description, and the input type is converted into the JSON schema the model sees. The scaffold includes a guide file for the developer's own coding agent, a runtime configuration file, a typed example tool, a smoke evaluation and a test that passes without an API key, so a project is testable from the first install. Agents run on LangGraph.js, and the developer keeps full access to the underlying graph rather than being locked into a proprietary runtime. B4.run is free and open source under the MIT licence. The site publishes no pricing and there is no hosted or paid tier; model costs depend on the provider the agent is configured to use. Because it builds on LangGraph.js, teams already using that library can adopt its file based conventions without replacing their existing execution model.