The LLM industry claims AI can write production code. We're here to test that — by building real products with an entirely AI-powered technical team. No developers. No designers. Just frontier models and domain experts.
Every product on this page was built without a single human engineer or designer. The entire technical stack — architecture, code, UI, deployment — is generated, reviewed, and operated by AI agents across frontier and open-weight models.
A local agent-neutral control plane. Self-hosted AI agent environment that manages context across models, tools, and workspaces — without vendor lock-in.
github.com/TDH-Labs/Harbor →An IPv6 VPN built on Yggdrasil mesh networking. Pure peer-to-peer — no centralized handshakes, no coordination server, no single point of control. Each node connects directly into the mesh.
Future ProductA domain expert identified the need. The AI team is scoping feasibility. If viable, it ships — no humans in the loop.
ResearchWe are not a traditional software company. We are an experiment running in public.
The builders and proponents of large language models say that AI can now write most if not all of the code required to ship production software. TDH Labs exists to test that claim — honestly, rigorously, and in public.
Our entire technical team is made up of AI agents. Frontier closed-weight models and open-weight models alike. They architect the systems, write every line of code, design every interface, manage every deployment, and triage every bug. No pull request is reviewed by a human engineer. No UI mockup is touched by a human designer.
The humans at TDH Labs are intentionally non-technical. They are domain experts — lawyers, accountants, operators, specialists in their fields — who have no training in software engineering, computer science, or design. They describe what the product needs to do. AI agents decide how to build it.
This is not a thought piece. We ship real software, for sale, under this constraint. Every product is a data point in an honest evaluation of what AI development can — and cannot — do when left to its own devices.
Fixed constraints that define the experiment. No exceptions.
Zero humans with software training touch the code. Full stop. AI agents own the entire technical stack.
The humans involved are specialists in law, finance, operations — not software. They define what, never how.
We use every available model — Claude, GPT, Gemini, DeepSeek, open-weight — selecting by capability, not loyalty.
Every product ships publicly. Every success and failure is a data point. We report honestly on what worked and what didn't.