UI-TARS
BetaMultimodal desktop + browser automation with vision-based selectors.
Multimodal desktop and browser automation stack with vision-based selectors, native OS integration, and persistent agent workflows. Strong fit for legacy-app automation.
- Category
- Browser & computer-use agents
- Runtime
- Python
- Pattern
- SDK / Library
- Adapter package
- intramedial-actor-uitars
Status
UI-TARS is Beta.
The integration path is built and tested. Customers can attach now; expect minor breaking changes until the 1.0 contract freeze.
Ships from the repo as `intramedial-actor-uitars` — not yet on a package registry.
How to attach
SDK / Library.
Install the adapter
Add intramedial-actor-uitars to your project.
Issue a workspace API key
Generate a key in the dashboard under Settings → API Keys, scoped to the workspace and environment this agent's data should land in.
Record each run
Construct the adapter once with your token and workspace id where you initialise your UI-TARS agent, then call recordRun(run) when a run finishes. The adapter maps the run to signed receipts and, by default, captures a failed attestation instead of crashing your agent.
Verify
Trigger one agent run. The Witness page shows the record once the agent has emitted it.
What gets recorded
2 event types, captured today.
- agent.invocation
- Every top-level agent run — the input, the resolved output, timing and token usage.
- tool.call
- Each tool or function the agent invokes, with its arguments and returned result.
Same category
Others in Browser & computer-use agents.
- browser-useBeta
Open-source Python browser agent. 50K+ GitHub stars.
- StagehandBeta
AI primitives on top of Playwright. TypeScript-first.
- SkyvernBeta
Vision-first browser agent. 21K+ stars, 10M+ executed workflows.
Vision-driven screen control. Available via API, Bedrock, Vertex.
Anthropic's official Chrome extension. Generally available.
- AgentQLBeta
Query language for AI agents extracting structured data from web pages.
Start recording UI-TARS agents.
Every UI-TARS run lands in the same canonical Witness record as every other framework in the catalogue: signed, replayable, and exportable as evidence.
A Trustvice ProductAgents act when the plane says yes, and the record says why.
Put admission in front of your MCP servers, repositories, SaaS tools and cloud accounts. Every agent gets a lease or a reason – and you keep the lever to stop any of them.
