CUGA

What is CUGA?

CUGA is an open-source harness for building agents, released by IBM Research under the Apache 2.0 license. The name stands for ConfigUrable Generalist Agent. You attach tools such as APIs, MCP servers or Python functions, add optional policies and company documents, and CUGA plans and carries out business requests. It can be embedded through a Python SDK, run from the CLI, or managed in a browser UI where you draft an agent, test it and publish a version. Policy types include Intent Guard, Playbook, Tool Approval, Tool Guide and Output Formatter, so a step like updating CRM records can pause for human approval. CugaSupervisor coordinates several specialist agents, and reusable SKILL.md files package procedures. CUGA runs locally, on Kubernetes with a Helm chart, or in air-gapped setups with your own models, and it supports OpenAI, watsonx, Azure OpenAI, Groq, OpenRouter, Gemini and Ollama. The software is free to use, so you pay only for models and hosting.

Use cases

• Automate business requests across apps
• Require approval before CRM writes
• Coordinate specialist agents
• Answer questions from company documents
• Run agents in air-gapped environments
• Embed an agent in a Python app

Who it's for

• Enterprise developers
• Automation engineers
• AI platform teams

Why use CUGA

• It runs on your infrastructure, so connectors, models and data can stay inside your boundary.
• Policies and approval gates let you decide which tool calls run automatically and which wait for a person.
• Tool support through MCP, OpenAPI and Python means existing internal services can be reused.
• The Apache 2.0 license allows use and modification without a license fee.
• Skills for Claude Code, Cursor, Codex and Bob help coding assistants install and build with CUGA.

CUGA features

• CugaAgent and CugaSupervisor for single and multi-agent work
• Tools from MCP, OpenAPI and Python functions
• Five policy types including Tool Approval and Intent Guard
• Knowledge ingestion with cited sources in answers
• Memory that feeds lessons from past runs into later ones
• Reusable SKILL.md agent skills
• Browser plus API execution modes
• Manage UI with draft and publish workflow
• Helm chart for Kubernetes deployment
• Multiple LLM providers via TOML or environment variables

CUGA integrations

• MCP
• OpenAPI
• OpenAI
• IBM watsonx
• Azure OpenAI
• Groq
• OpenRouter
• Google Gemini
• Ollama
• Langfuse
• Kubernetes
• LangGraph

How to get started with CUGA

1. Install CUGA and choose a path: run the demo, use the Python SDK, or start the Manage UI with cuga start manager.
2. Connect a model provider through a .env file and settings TOML, then attach tools, policies and documents.
3. Send a request; the orchestrator analyzes it, decomposes the work, routes it and executes the steps.
4. Policies can pause selected actions for human approval before they run.
5. Publish the agent version and deploy it locally, with the Helm chart on Kubernetes, or in an air-gapped setup.

CUGA limitations

• You host it yourself and supply your own model access
• Air-gapped use needs your own reasoning model and inference server
• Intent matching and guidance depend on how policies are configured

CUGA pricing

CUGA is open source, so you can use and self-host it for free.

Prices can change and may vary by country or exclude tax. Check the official pricing page.

CUGA alternatives

The main CUGA alternatives listed on Debuly are LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Agno and Amazon Bedrock Agents. The table compares pricing and platforms; order is A–Z.

1. 

   Agno

   Python framework and AgentOS runtime for building agents, teams and workflows, then serving them as an API with a control plane.

   Developer platformsAutomation and agentsCoding
   Open sourceVisit Agno (official site)
2. 

   Amazon Bedrock Agents

   AWS service for building agents that use foundation models, company APIs and data to complete multistep tasks, with memory and multi-agent teams.

   Automation and agentsDeveloper platforms
   PaidVisit Amazon Bedrock Agents (official site)
3. 

   AutoGen

   Microsoft's MIT-licensed Python and .NET framework for multi-agent AI apps, now archived; new projects are pointed to Microsoft Agent Framework.

   Developer platformsAutomation and agentsCoding
   Open sourceVisit AutoGen (official site)
4. 

   CrewAI

   Multi-agent framework and platform for building AI agent crews in Python or a visual editor, then deploying and monitoring them.

   Developer platformsAutomation and agentsCoding
   Open sourceVisit CrewAI (official site)
5. 

   LangGraph

   Low-level open-source orchestration framework and runtime for building stateful, long-running AI agents as graphs.

   Developer platformsAutomation and agentsCoding
   Open sourceVisit LangGraph (official site)
6. 

   OpenAI Agents SDK

   Open-source Python framework for building multi-agent workflows with tools, handoffs, guardrails, tracing, sandboxes and voice agents.

   Developer platformsAutomation and agentsCoding
   Open sourceVisit OpenAI Agents SDK (official site)

CUGA FAQ

Sources & verification

Facts on this page come from the official sources below, each with the date we read it. Written by Debuly; the vendor did not review this page.

Image credits

• Logo: from cuga.dev · fetched 23 Oct 2026 · Vendor trademark, used only to identify the product; removed on request
• Cover: none collected yet, so a generated tile is shown.

Spot an error? Report an issue with this listing. We also remove a logo or image on the owner's request.