Start
System Components
Every piece of CUGA mapped to the five architectural layers.
CUGA is organized into five layers. This table maps each product piece to its layer and where to read more.
| Piece | What it does | Docs |
|---|
| Web UI | Chat and Manage UI in the browser | Demos · Manage |
| Python SDK | CugaAgent / CugaSupervisor in your app | CugaAgent |
| CLI | cuga start, knowledge/policy helpers | CLI presets |
| Piece | What it does | Docs |
|---|
| CugaAgent | Single-agent graph — plan, execute, answer | CugaAgent |
| CugaSupervisor | Router across several agents (local or A2A) | CugaSupervisor |
| Policies | Intent Guard, Playbook, Tool Guide, Tool Approval, Output Formatter, ToolGuard | Policies |
| Human-in-the-loop | Suggest action → wait for approval → continue | Tool Approval |
| Reflection | Second look after a step; retry if off-track | Todos, reflection, shortlisting |
| Todos | Checklist for multi-step CugaLite tasks | Todos, reflection, shortlisting |
| Tool shortlister | find_tools when the catalog is large | Todos, reflection, shortlisting |
| Context summarization | Compress older turns in long threads | Context summarization |
| Agent spawn | Nested CugaAgent mid-run (spawn_agent) | How it works — Multi-agent |
Internal graph nodes (Task Analyzer, Plan Decomposition, Browser/API Planners, Code Agent, Action Agent) run automatically — you configure agents and policies, not individual nodes.
| Piece | What it does | Docs |
|---|
| LLM Client | Provider abstraction (OpenAI, WatsonX, Groq, …) | Models |
| Browser Runtime | Playwright automation for web tasks | Task modes |
| Tool Adapter | Python tools, OpenAPI, MCP | Tools |
| Knowledge Engine | Docling ingest, embeddings, search, citations | Knowledge |
| Memory | Learn from past runs (guidelines + trajectories) | Memory |
| Agent Skills | SKILL.md playbooks via load_skill | Skills |
| Sandbox | Code execution backends (native, e2b, …) | Sandbox |
| Piece | What it does | Docs |
|---|
| Vector DB | Milvus or pgvector for knowledge/policy vectors | Storage |
| Config DB | Conversations, policies, published agent versions | Storage |
| Piece | What it does | Docs |
|---|
| LLM providers | Remote model APIs | Models |
| MCP / OpenAPI servers | External tool sources | Tools |
| Secrets vault | API keys and credentials | Secrets vault |
| Observability | OpenLit, Langfuse, Activity tracker + cuga-viz | Observability |
Full walkthrough: How CUGA works.
