finds.dev

cuga-project/cuga-agent

CUGA is an open-source generalist agent harness for the enterprise, supporting complex task execution on web and APIs, OpenAPI/MCP integrations, composable architecture, reasoning modes, and policy-aware features.

View on GitHubHomepage ↗

★ 1,061PythonApache-2.0updated Nov 2026

// the take

CUGA is a self-hosted agent harness aimed at enterprises that want LLM agents acting on internal APIs and web UIs without ceding control to a hosted platform. It bundles policy enforcement (approval gates, intent blocking, tool guards), OpenAPI/MCP tool wiring, and multi-agent orchestration via LangGraph, targeting teams that need an auditable, air-gappable alternative to something like an OpenAI/Anthropic agent SDK wrapper.

The policy layer is actually enforced, not just prompted: ToolGuard code runs before the underlying tool call and can hard-block it, which is a real gate rather than instructions the model can ignore. Run receipts give per-invocation token/call/latency accounting out of the box, which most agent frameworks leave to you to bolt on. It supports both OpenAPI and MCP as first-class tool sources plus a supervisor pattern for routing to specialist sub-agents, so you're not locked into one tool protocol. The air-gapped deployment path (UBI image + model preloader script) suggests someone actually tested disconnected enterprise environments rather than just claiming support for them.

The feature set advertised in the README (memory/Evolve, scheduled events, citations) mostly lives in separate services you have to deploy and wire up yourself (a distinct events service, a Postgres-backed Evolve backend) — the core harness alone doesn't give you most of what's shown in the screenshots. The benchmark numbers (AppWorld, WebArena) are self-reported by IBM with a specific model (GLM-5.3-Flash) and no independent reproduction, so they tell you little about behavior with whatever model you actually plan to run. At ~1k stars and 203 forks this is early for something pitching itself as enterprise infrastructure with OIDC auth, Kubernetes Helm charts, and policy compliance claims — there's not much evidence yet of production battle-testing outside IBM's own use. The surface area (browser automation + API tools + multi-agent + events + memory + guardrails) is large enough that getting all the pieces configured correctly (ToolGuard code, approval flows, OIDC roles) is its own project before you ship anything.

We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
1,061
Forks
203
Language
Python
License
Apache-2.0
Last commit
2 Nov 2026

// nearby

If cuga-agent isn't quite it.

All repos →