Henry Mao

Henry Mao built Smithery, an open MCP registry and gateway for AI agents, after co-founding Jenni, an AI academic writing copilot. He also publishes research as Huanru Henry Mao, including a sole-authored paper proposing decaying fast weights as an alternative to language-model attention.[2][3]

After leaving Jenni in 2024, Mao built Smithery and scaled its infrastructure. On his personal website, he reports that it processed 100,000 tool calls from agents per day; he does not specify the measurement period. Smithery provided a registry and hosting platform designed to make finding and running an MCP server a one-click operation.[2][4]

Mao hosted the MCP track at the 2025 AI Engineer World's Fair, receiving acknowledgment in Latent.Space's conference recap. On August 5, 2026, Arcade.dev announced its acquisition of Smithery, with Smithery co-founder Anirudh Kamath joining the Arcade team.[1][4]

Before his work on agent infrastructure, Mao studied neural networks for music and natural language generation. He entered UC San Diego's PhD program in 2018 and left two years later to co-found Jenni. He exited Jenni in 2024. This earlier period included both generative-model research and building an AI writing product.[2]

His research publications include DeepJ: Style-Specific Music Generation (2018), with Taylor Shin and Garrison W. Cottrell, and Improving Neural Story Generation by Targeted Common Sense Grounding (2019), with Bodhisattwa Prasad Majumder, Julian McAuley and Cottrell. For ReZero is All You Need: Fast Convergence at Large Depth, his publication list credits Thomas Bachlechner, Majumder and Mao with equal contributions, alongside Cottrell and McAuley; it lists the work as 2020 research appearing in the UAI 2021 proceedings.[2]

In Fine-Tuning Pre-trained Transformers into Decaying Fast Weights, submitted on October 9, 2022, Mao addressed the cost of autoregressive language generation: self-attention incurs O(T) computation per generated token as the sequence grows. The paper examines recurrent approximations intended to achieve constant time and memory complexity, then proposes decaying fast weights as a simpler alternative. Its abstract reports fast GPU execution, competitive performance on WikiText-103 and retention of “99% of attention's performance for GPT-2,” without defining that percentage's metric in the abstract.[3]