Quick Start
CUGA — configurable generalist agent for enterprise automation. Pick a path: try the demo, use the SDK, or manage an agent in the UI.
Quick Start
CUGA (ConfigUrable Generalist Agent) is an open-source agent harness from IBM Research. You attach tools (APIs, MCP, Python functions), optional policies and documents, and CUGA runs the task — in your app via the SDK, or as a server with cuga start.
Not sure where to begin? Pick one path:
Try the demo
cuga start demo — chat UI with sample tools in under 5 minutes
Use the SDK
CugaAgent(tools=[...]) → await agent.invoke(...) in Python
Manage an agent
cuga start manager — draft tools & policies in the UI, then publish
Install first if you have not already: Installation.
Want the mental model before diving in? Read How CUGA works.
CUGA is built in five layers — Clients → Orchestrator → Capabilities → Persistence → External services. Your path above maps to the Clients layer; the orchestrator and capabilities run behind it.
Benchmarks
What you get
- CugaAgent — one agent, your tools,
invoke()/stream() - CugaSupervisor — delegate across several agents (local or A2A)
- Policies — Intent Guard, Playbooks, Tool Approval, ToolGuard, and more
- Knowledge — ingest PDFs/Office/HTML, cite sources in answers
- Memory — learn from past runs; inject guidelines next time
- Agent skills — reusable
SKILL.mdplaybooks - Manage & self-host — draft in the browser, publish a version, deploy with Helm
Full component map: System components.
Ecosystem
cuga-agent is the harness. These repos sit beside it:
| Repo | Role |
|---|---|
| cuga-skills | Catalog of runtime SKILL.md playbooks you install into .cuga/skills |
| cuga-harness-kit | IDE skills that teach Cursor, Claude Code, Codex, and Bob how to install and build with CUGA |
| cuga-eval | Evaluation harness for running agent benchmarks against CUGA |
Skills in the product: Build → Skills. Benchmarks: cuga-eval.