Playground

Type an instruction and watch how the AI orchestrates operations within the capability-manifest whitelist — producing an operation plan and highlighting the matched capabilities.

① Operation plan produced by the LLM

Switching to the leave page and filling in the request

1. navigate → leave
2. fill → leave.type = "Personal"
3. fill → leave.days = "1"
4. fill → leave.reason = "Family matter"
5. click → leave.submit

② Matched capabilities (whitelist)

leave · Leave

action: leave.submit

field: leave.type (select)

field: leave.days (number)

field: leave.reason (text)

employees · Employees

action: employees.add

field: employees.name (text)

field: employees.dept (select)

This is a static demo: preset manifest + mock provider, no real LLM. In production the LLM orchestrates within the manifest whitelist, and any target outside the list is rejected by parsePlan.

What it demonstrates

• Read structure, don't guess the screen: the capability manifest on the right is extracted from the code structure at build time. The LLM can only orchestrate within this whitelist.
• Precise orchestration: one natural-language sentence → a deterministic sequence of navigate / fill / click steps, each corresponding to one capability in the manifest.
• Whitelist constraint: modules or operations outside the manifest are rejected by parsePlan and never executed.

This page is a static demo (preset manifest + mock provider, no real LLM). For real integration see Quick Start.