ByteDanceTARS is a multimodal AI agent stack from ByteDance that currently ships two projects: Agent TARS and UI-TARS Desktop. Both bring graphical-user-interface (GUI) agents and vision to real-world tasks, using multimodal LLMs to complete work across a terminal, computer, and browser.
Agent TARS is a general multimodal agent stack that ships as a command-line interface and a web UI, in headful or headless server modes. It controls a browser through a GUI agent, the DOM, or a hybrid of the two, and its kernel is built on the Model Context Protocol (MCP) so external tool servers can be mounted. A protocol-driven event stream carries context between steps and drives the agent UI.
UI-TARS Desktop is a native GUI agent for a local computer, driven by the UI-TARS model and the Seed-1.5-VL/1.6 series. It operates a local or remote computer as well as a browser, with remote computer and browser operators available without configuration. A separate UI-TARS SDK provides a cross-platform toolkit for building GUI automation agents.
Features
- Two projects: Agent TARS for general multimodal task completion and UI-TARS Desktop for native GUI operation
- GUI and vision agents: vision-language models locate and act on on-screen elements to drive applications
- Hybrid browser control: GUI-agent visual grounding, DOM access, or a combined strategy
- Local and remote operators: control a local computer, or a remote computer and browser without setup
- MCP kernel: mount external MCP tool servers to reach real-world tools
- CLI and web UI: run Agent TARS headful with a web UI or headless as a server, launched via npx or a global install
- Provider choice: run against providers such as Volcengine or Anthropic by passing a model and API key
- UI-TARS SDK: a cross-platform toolkit for building GUI automation agents
