Research/AI
How fast can an agent read a screen?
We timed every open computer-use server we could run reading the same window on Windows, macOS and Linux. Munim Computer Use was the fastest at every folder size on all three systems and returned the fewest tokens.
Nov 4, 2026 · 5 min read

A computer-use agent spends most of its time looking. Before every click it has to read the window in front of it: which buttons exist, what the list contains, where the text field is. If reading is slow or verbose, every step of every task pays for it, in seconds and in tokens.
So we measured that step directly. On 4 November 2026 we ran the latest release of every open computer-use server we could install against the same file manager window on each operating system, while the folder grew from 100 to 20,000 files. Munim Computer Use was the fastest at every folder size on Windows, macOS and Linux, and returned the fewest tokens on all three.
0.12 s
per read of File Explorer on Windows, flat from 100 to 20,000 files
0.10 s
per read of a 20,000-file Finder window on macOS
1.0–2.0K
tokens per read, the fewest on every system
The test
Each server read a file manager window with its own read tool: File Explorer on Windows 11, Finder on macOS 27 and Thunar on Linux. Where a tool could return a screenshot alongside the text, we turned the screenshot off so every server was compared on text. For each folder size a fresh window was opened, every server was started once and called over stdio, one warm-up call was dropped, and five calls were timed. The whole run was done twice, so each time below is the median of ten calls. Tokens are estimated at four characters per token.
Windows 11 · File Explorer, 20,000 files: seconds per read
macOS 27 · Finder, 20,000 files: seconds per read
Linux · Thunar, 20,000 files: seconds per read
Fast is not enough: reading the right rows
Speed only matters if the read is useful. A good read lists what is on screen and nothing else, so the agent can act on it and the context window stays small. On Linux only Munim Computer Use listed exactly the 33 rows in view: one server listed files only at the smallest folder size, another listed none, and a third listed every row it had, on screen or not, which is why its reads ran to 26,000–43,000 tokens.
Tokens per read (≈, 4 characters per token)
| Server | Windows | macOS | Linux |
|---|---|---|---|
| Munim Computer Use 0.10.0 | 2.0K | 1.0K | 1.2K |
| Cua Driver 0.38.0 | 3.9K | 2.4K | 2.3–5.7K |
| computer-use-mcp 7.8.0 | 9.3–9.4K | 22.4K | 28.6K |
| open-computer-use 1.1.3 | 4.9K | 1.1K | 26.0–42.7K |
| Windows-MCP 0.8.8 | 1.2–3.9K | — | — |
| MacOS-MCP 0.4.7 | — | 2.6–17.3K | — |
Very large folders exposed a limit in the Linux accessibility stack itself: AT-SPI refuses to hand over more than 65,536 children in one request, so release 0.9.3 read nothing of the 20,000-file folder. Release 0.10.0 walks such a list by index and reads it in 0.23 s.
Getting faster release by release
We also read one File Explorer window with every Windows release, from the release binaries, on 27 October 2026. Two changes account for the gains: 0.7.0 started reading only what is on screen and fetching each element's attributes in one request, and 0.8.0 fetches the whole window in one request instead of one per element, with the same output.
Munim Computer Use on Windows: seconds per read, by release
Acting
On the same Windows PC we also timed acting: the five clicks of 1 + 2 = on Calculator, each a separate call by element id after one read. Munim Computer Use took 1.70 s (0.34 s per click) against 2.31 s for Cua Driver, median of three runs. Most of a click is the visible agent pointer flying to its target, so the user can see where the agent is acting; run_steps sends all five clicks in one call.


