Therry Mirandau2-agent
LLM-native Android automation, 85% fewer tokens
u2-agent is an LLM-native Android automation tool built on Bun and TypeScript. It talks to a native uiautomator2 JSON-RPC daemon over ADB and deduplicates the view hierarchy into a compact semantic tree where every actionable element gets an ephemeral handle (@1, @2, …). Built for the OpenCode Go ecosystem and shipped with CI, docs and a compiled single-file binary.
- Role
- Creator & Engineer
- Client
- Own product
- Category
- AI / Mobile Automation
- Headline figure
- ~250
u2-agent — HeroVerified
- tokens per screen (vs ~15,000)
- ~250
- cached action latency
- <15ms
- fewer tokens
- 85%
Problem
Existing mobile automation floods the context window with 5,000–20,000 tokens per snapshot, causing high cost, lost-in-the-middle reasoning failures and 500–2000ms latency per step.
Solution
An in-memory daemon and a semantic tree with ephemeral handles, collapsing a screen to ~250 tokens and cached actions to under 15ms — a zero-dependency CLI with no framework overhead.
Capabilities shipped
Semantic tree
Deduplicated view hierarchy with ephemeral handles (@1, @2, …).
Sub-15ms actions
In-memory daemon and cached actions instead of XML serialization.
Zero dependencies
Built on Bun and TypeScript native APIs, shipped as one binary.
3-step agent loop
A minimal perceive → decide → act cycle designed for LLMs.
Developed for agents
Output shaped for context efficiency and reasoning accuracy.
Stack
- Runtime
- Bun
- Language
- TypeScript
- Integration
- ADBuiautomator2
- Deployment
- Cloudflare Pages
