Flue has launched as a TypeScript framework for building autonomous agents that move beyond simple chatbot functionality. Unlike traditional SDKs, Flue provides a programmable harness architecture similar to what powers Claude Code and Codex. Developers write agents in TypeScript and deploy them anywhere—as HTTP servers or CLI tools for local development. This flexibility allows teams to choose the deployment model that best fits their workflow.

The framework gives developers complete control over the agent lifecycle: skills as reusable workflows, session-based memory tracking, shell execution capabilities, and fine-grained security for API tokens. Flue includes a built-in zero-config virtual sandbox, though users can connect their own remote sandbox if preferred. An example demonstrates building an AI issue triage system in just 22 lines of TypeScript code.

Flue positions itself against off-the-shelf AI tools like Dosu, Greptile, and CodeRabbit, arguing that these general-purpose solutions struggle to fit specific products, data, or workflows. The framework gives developers ownership of the entire stack—the model, harness, and sandbox—enabling them to build agents tailored to their exact needs rather than renting generic solutions. This represents a significant shift from renting AI capabilities to owning them.