oriel Comparisons Updated

oriel and Cua Driver compared

Both are MCP servers that read the Windows accessibility tree. On WindowsAgentArena round 9 (October 2026), with the same model, tasks took a median 49 s with oriel and 70 s with Cua Driver 0.38.0, at 38% lower model cost. Cua Driver, without 4 of its 63 tools, completed 4 points more. If completion is what you need most, Cua Driver is a fair choice.

At a glance

orielCua Driver
How it sees a windowThe accessibility tree, as textThe accessibility tree, as text, with screenshots on request
What an action changedIn the same replyA second call, get_window_state
Several steps in one callbatchrun_steps, and run_script for JavaScript over its tools
Tools1963 on Windows (0.38.0)
LicenceApache-2.0MIT

Whole tasks: WindowsAgentArena round 9

WindowsAgentArena is Microsoft's benchmark of everyday tasks in a Windows 11 VM, scored by its own scripts. Both layers ran Claude Sonnet 5.5 in Claude Code, with the same prompt and a 50-turn limit; the prompt differs in one sentence, which names the layer's tools. Two runs each on fresh VMs, October 2026, 153 tasks scored in all four runs.

oriel v12Cua Driver 0.38.0
Time per task, median49 s70 s
Time per task, mean74 s90 s
Model cost per run of 153 tasks$38$61
Model turns per task21.827.0
Tasks completed, mean of 2 runs96.2 (63%)102.7 (67%)
Each run96 / 9798 / 108
LibreOffice Writer, 19 tasks5.08.9

What the numbers show, and what they don't

  • Time, cost and turns. With oriel, the median time was 30% lower, the model cost 38% lower and there were 19% fewer turns. Each gap is larger than the gap between either layer's own two runs. Swapping the server did this; the benchmark does not show which feature did it.
  • Not only failing sooner. On the 77 tasks that all four runs completed, oriel took a median 46 s against 63 s, with 17.8 against 23.7 turns.
  • Completion. Cua Driver completed 4.2 points more (67% against 63%). Its own two identical runs differed by 6.5 points, so that gap is within run-to-run variation. But its lower run (98) still beat oriel's higher run (97), and it completed more in every round since round 5. Most of the gap is LibreOffice Writer.
  • Earlier rounds. The time gap appeared in round 8. In rounds 1 to 7, oriel's median time was the same as Cua Driver's or longer. oriel's model cost was about a third lower in every round.
  • Hidden tools. Cua Driver ran without 4 of its 63 tools. run_script runs JavaScript over Cua Driver's own tools in a sandbox with no file, network or process access; it is on by default. It was hidden because oriel has no script tool, and with it Cua Driver might need fewer turns, less time and less cost. install_ffmpeg and install_extension install software, and check_for_update checks GitHub for a newer driver. Cua Driver with run_script was not measured.
  • Version. Round 9 ran Cua Driver 0.38.0, released 21 Oct 2026. The newest on 10 Nov 2026 is 0.46.1, which has not run on WindowsAgentArena.
  • Cost is the model cost Claude Code reported for each task, a list-price estimate from the token counts, not a bill. Both layers are priced with the same table.

Every round, the intervals and the raw rows: What this does not show, WindowsAgentArena in docs/benchmarks.md, bench/waa/RESULTS.md.

Per call: 10 Nov 2026, Cua Driver 0.46.1

The same tasks through both servers on one unlocked Windows 11 PC, driven from a MacBook on the LAN. oriel 0.1.0 against Cua Driver 0.46.1, the newest release that day, through its daemon. 5 runs per task, median wall time per MCP call.

Taskoriel 0.1.0Cua Driver 0.46.1
List windows1 ms30 ms
Launch Character Map192 ms604 ms
Launch and return its tree294 ms, 1 call2 calls
Read the Character Map window20 ms37 ms
Fill a field2 ms3 ms
Press a button2 ms5 ms
Press and report what changed42 ms, 1 call2 calls
Correct outcome10 of 105 of 5
  • To read a window, Cua Driver first lists windows to find it: 30 ms more.
  • Cua Driver 0.38.0, run the same day, was within 15 ms of 0.46.1 on every row. The larger gaps measured on 21 Oct 2026 came from that day's conditions, most likely the locked PC. That is an inference: the open windows differed too.
  • Bare cua-driver mcp shows its agent cursor by default, which adds about 565 ms to each fill and press. With the cursor off it matches the daemon within 2 ms.

Rows and method: bench/headtohead/RESULTS.md.

Which to choose

  • Cua Driver if completion is what you need most, if most of your work is in LibreOffice Writer, or if you want a script tool in the server.
  • oriel if you want each task done in less time and fewer turns, the change in the same reply, and one exe under 200 KB with no admin rights, which keeps password fields, sign-in, UAC prompts, password managers and terminals off limits by default.

Questions

Did you test an old Cua Driver?

WindowsAgentArena round 9 ran Cua Driver 0.38.0 (October 2026). The per-call run of 10 Nov 2026 used 0.46.1: oriel took less time on every call, by less than in October, and 0.38.0, run the same day, was within 15 ms of 0.46.1. No WindowsAgentArena round has run 0.46.1.

Why was run_script hidden?

oriel has no script tool, so scripting was kept out of the comparison. run_script runs JavaScript over Cua Driver's own tools in a sandbox with no file, network or process access. It is not a shell. Cua Driver with run_script was not measured.

Is Cua Driver's completion lead real?

Cua Driver 0.38.0 completed 4.2 points more (67% against 63%). Its own two identical runs differed by 6.5 points, so the gap is within run-to-run variation. But Cua Driver completed more in every round since round 5.

Try oriel

Free and open source, one exe under 200 KB, no admin rights. Tested with Claude Code on Windows 11.

Download for Windows Set-up steps

Not code-signed yet, so Windows will warn. Why, and how to check the file

Sources