Henry Mao is the co-founder and chief executive of Smithery, which builds infrastructure connecting AI agents to external software and services. His work addresses a practical limitation of powerful language models: intelligence alone cannot complete useful tasks without trustworthy access to tools, context, and real-world systems.
From generative-model research to Jenni AI
Mao studied computer science at the University of California, San Diego, earned a master’s degree, and entered its doctoral program in 2018. Publishing as Huanru Henry Mao, he co-authored DeepJ, a neural model for generating music in controllable combinations of compositional styles, and subsequently investigated multi-instrument music generation and commonsense grounding for generated stories.
With David Park, whom he met in college, Mao explored several unsuccessful startup ideas before co-founding Jenni AI and becoming its chief technology officer. He left his doctoral program to build the company full time. Originally developed for marketing content, Jenni pivoted toward academic writing after its founders discovered students were adopting the product independently. Mao concentrated on features students needed while Park developed distribution, an evolution he describes in his account of their early experiments and product pivot. Mao exited the company in 2024.
Making agents operational
Experiments with ARC-AGI led Mao to a sharper distinction between benchmark intelligence and practical autonomy: solving abstract reasoning problems does not equip a model to inspect a repository, authenticate with another application, or finish a workflow. He founded Smithery around December 2024 to close that gap and participated in the Model Context Protocol steering committee.
Discovery requires trust. A registry can list servers without establishing their quality, security, or suitability for a particular task.
Agent experience determines usability. Installation, credentials, stateful sessions, resumability, hosting, and observability shape whether software can use a service reliably.
Agentic payments need workable economics. Developers require distribution and compensation models that do not force users into separate subscriptions for every specialized tool.
Orchestration turns integrations into completed work. Mao demonstrated an agent locating a GitHub issue, connecting to Linear, and creating a corresponding ticket across two independent services.
For Mao, useful agents depend on an ecosystem whose capabilities are discoverable, dependable, economically sustainable, and designed for software to act on a person’s behalf.
