cua
Facts
- Openness
- Public code, licence not stated
- Language
- Python
- Pinned version
0637a56(2026-10-01)- Topics checked
- 28 of 33
- Kind of agent
- Browser/GUI
What it does notably well
- Budgets
cua's BudgetManagerCallback is a USD-only budget (sum of response_cost from on_usage) scoped to one run by default (reset_after_each_run=True), off unless max_trajectory_budget…
- Compaction
cua's only history-shortening step is ImageRetentionCallback, a count-based prune: when only_n_most_recent_images is set it keeps the newest N screenshot outputs and removes each…
- Concurrency
cua's ComputerAgent has an async core (run is an async generator, model calls and handlers are awaited, sync function tools run via asyncio.to_thread) in which tool calls of one…
- Config and routing
cua selects its agent loop by model string through an @register_agent regex-priority registry: 20 @register_agent registrations in cua_agent/loops (priorities from -100 for the…
- Errors and failover
cua retries transient model-call errors with exponential backoff and random jitter (up to max_retries, default 3) and feeds a ToolError raised by a computer action, unknown…
- Eval harness
cua has a gym-style evaluation environment in cua-bench (make, reset, step, evaluate), benchmark adapters for OSWorld and GUI grounding sets, a HUD integration, a…
See how cua does on each topic
The topic-by-topic verdicts, the code lines and a side-by-side comparison are in the app. Sign in with Google; new accounts are approved by hand.
Independent analysis, not affiliated with cua.