How turixai/turix-cua works
TuriX is a desktop computer-use agent: a user gives it a task, planning and model reasoning produce structured actions, and the controller executes them against the desktop. The tree also supports planning-time search, skill guidance, and persistent records used for memory or requested file reads. The README documents broader CLI, work/code, and MCP workflows, but this checkout’s sampled implementation centers on the macOS agent/controller; those broader capabilities are not mapped as separate runtime components because their wiring is not evidenced here.
Main entry points: agent/service.py Agent runtime · planner_service.py Task planner · controller/service.py Action controller · brain_search.py Read-and-reason flow
Components 17
Planning and reasoning
- Agent runtime
src/agent/service.py - Task planner
src/agent/planner_service.py - Message manager
src/agent/message_manager/service.py - Output schemas
src/agent/structured_llm.py - Agent prompts
src/agent/prompts.py
Desktop execution
- Action controller
src/controller/service.py - Action registry
src/controller/registry/service.py - UI tree builder
src/mac/tree.py - UI elements
src/mac/element.py - Mac input actions
src/mac/actions.py
Context and memory
- Read-and-reason flow
src/utils/brain_search.py - Record store
src/utils/record_store.py - Skill guides
src/utils/skills.py
External systems
- Configured language model
- macOS desktop
- Web search
Other components
- Task requester
Connections 21
- Task requester → Agent runtime: submits task
- Agent runtime → Task planner: plans task
- Task planner → Configured language model: requests planning
- Task planner → Web search: searches web
- Task planner → Skill guides: selects guides
- Task planner → Agent runtime: returns plan
- Agent runtime → Message manager: builds context
- Message manager → Agent prompts: uses prompts
- Agent runtime → Configured language model: requests actions
- Configured language model → Output schemas: structured output
- Agent runtime → Action controller: dispatches actions
- Action controller → Action registry: executes action
- Action controller → UI tree builder: builds UI tree
- UI tree builder → UI elements: creates nodes
- Action controller → Mac input actions: performs input
- Mac input actions → macOS desktop: sends events
- macOS desktop → UI tree builder: exposes UI state
- Agent runtime → Read-and-reason flow: handles read request
- Read-and-reason flow → Record store: reads records
- Read-and-reason flow → Message manager: updates context
- Agent runtime → Record store: stores memory