Zhengyi Luo built AgentsDock to run AI research agents from his phone

Why it matters

AgentsDock addresses the operational work around coding agents: supervising jobs, remote machines and research artifacts after the prompt is sent.

Zhengyi "Zen" Luo, a senior research scientist at Nvidia, is testing AgentsDock, a desktop and mobile workspace for supervising coding agents across the computers researchers already use. The current desktop build is version 0.2.13-beta.33, with clients for macOS, Windows and Linux alongside beta access for iPhone, iPad and Android.

Luo originally built the system to orchestrate robotics-training jobs and now runs much of his robotics work from his phone.

That origin explains why AgentsDock looks different from a general-purpose AI code editor. Its interface gives plots, videos, files and rendered experiment rollouts the same billing as chat and source code. Researchers can inspect a training result on mobile, reconnect to a persistent terminal and send another instruction without returning to the machine running the job.

Luo arrived at the problem through a decade of computer vision and robotics research. He earned a bachelor's degree in computer science from the University of Pennsylvania and a doctorate at Carnegie Mellon University's Robotics Institute. His Ph.D. work, advised by Kris Kitani, focused on universal humanoid control. The supplied research record describes an AR/VR software engineering internship, a research-scientist internship at the University of Toronto's AI Lab and a visiting researcher role at Meta Reality Labs, followed later by work at NVIDIA. Luo's official website lists him as a research scientist at Nvidia's GEAR Lab, while his public CV details his education, publications and research roles.

A control surface above the agents

AgentsDock currently supports Anthropic's Claude Code, OpenAI's Codex and Cursor, the AI code editor. It leaves model execution and account management to those providers. Users install and authenticate the relevant command-line tools on their own machines, then connect AgentsDock as the interface for starting chats, following progress and reviewing outputs.

The architecture has two pieces. The client runs on a computer, phone or tablet. The public AgentsServer repository runs on a Mac or Linux machine controlled by the user and provides access to agent sessions, files, scheduled jobs, media and persistent tmux terminals. A researcher can connect a lab workstation, a Mac mini at home and a rented GPU box, then switch among them from the same app.

AgentsDock's setup documentation says model requests travel through the provider tools installed on the user's machine rather than being proxied through AgentsDock. Users authenticate Claude Code or Codex on the remote machine. That design gives research groups a route to use several commercial agents while keeping project files and compute on hardware they control.

The AgentsDock homepage describes remote code viewing and editing plus inline plots, images and rendered rollouts. The AgentsServer repository documents persistent chats, queues, scheduled jobs, files and videos. AgentsDock's homepage demo shows a humanoid locomotion and manipulation run with 92.4% success, 0.18-meter final error and 42,000 steps. Those figures describe the sample experiment, not AgentsDock performance or customer results.

Built from a research workflow

The strongest part of Luo's pitch is its specificity. Coding agents are usually presented through software-engineering tasks that end with a pull request. Robotics research produces a messier trail: checkpoints, logs, simulation videos, plots, terminal processes and jobs spread across several machines. AgentsDock treats those artifacts as part of the agent conversation.

That also clarifies where AgentsDock fits beside the tools it supports. Claude Code, Codex and Cursor supply the agent capabilities. AgentsDock handles the operational layer around them: which machine is running a session, whether the process is still alive, what output it produced and how a researcher can intervene from another device.

Provider competition works in Luo's favor. Researchers can use several agents in one workspace while keeping projects and compute on their own machines. AgentsDock does not need to build a foundation model or persuade users to abandon existing provider accounts. Its job is to make a growing collection of agents manageable.

Several adjacent products are pursuing parts of the same job. Nimbalyst, an open-source local-first workspace, centers on coding agents. Conductor focuses on running teams of coding agents on the Mac, while Aiola offers an open-source command center for multiple providers. Cognition's Devin Desktop also brings its coding agent onto local computers. AgentsDock's distinction is its combination of mobile access, multiple remote servers, cross-platform clients and support for research outputs such as plots and rollout videos.

Capital is moving into the surrounding category. Cursor announced a $2.3 billion Series D at a $29.3 billion valuation in November 2025. Niteshift, another AI coding company, raised a $7 million seed round in June 2026. Those companies pursue broader coding workflows, but their funding shows the resources available to vendors competing for the interface around AI-assisted software work.

AgentsDock's homepage says it is "trusted by researchers at" Carnegie Mellon University, UC Berkeley and Nvidia. The site does not name institutional contracts or deployments. A supplied snapshot of the AgentsServer repository recorded 55 stars, three forks, 539 commits, 81 branches and 157 tags, figures that can change with repository activity.

No legal entity, headquarters, headcount or formal executive team has been publicly established for AgentsDock. Luo's public profiles place him in the San Francisco Bay Area. The product has disclosed no user count, customer count, revenue, usage volume or pricing, and no funding round, valuation or named investor is publicly identified. Its download page currently offers beta builds without a disclosed commercial plan.

Self-hosting comes with an operator manual

Keeping code and credentials on user-controlled hardware reduces one concern and creates another: users become responsible for securing the server that can reach their files, terminals and authenticated coding agents.

The AgentsServer repository warns users against exposing port 7850 directly to the public internet. That warning matters because AgentsDock can provide broad access to the same machine that holds research code, experiment data and provider credentials. Tailscale is offered as an optional private-network layer, and tmux can keep terminal sessions alive between client connections.

The beta supports desktop clients for macOS, Windows and Linux, with mobile distribution through TestFlight on Apple devices and a downloadable APK on Android, according to the product homepage and release listings. The homepage specifies macOS 14 or later, Windows 10 or later, Linux builds for x86_64 and ARM64, iOS and iPadOS 17 or later, and an ARM64 Android beta. The listed Android build is version 0.1.1-beta.7.

Luo's bet is that the interface around agents will become its own product category as researchers hand off jobs to several models on several machines. His advantage is direct exposure to the work AgentsDock is designed to supervise. The product grew from Luo's effort to orchestrate robotics-training jobs and monitor them from his phone.

AgentsDock remains an early beta with unanswered distribution and business questions. Its product thesis is already clear: agents working across a collection of computers need an interface that covers more than chat. Luo would like that control room to fit on a phone.