Mel – the agentic terminal
The AI terminal for teams.
Benchmarks
Mel v0.1.62 on Auto Balanced passed 46 of the 89 tasks in Terminal-Bench 2.0, with one try per task. That is 52%. This is our own run, not a score on the official leaderboard.
Self-reported. Run on October 24 and 25, 2026.
The result
Mel passed 46 of the 89 tasks. With one try per task, the 95% range is 41% to 62%.
Each task ran once. A task counts as passed only when the task's own tests passed.
The setup
Terminal-Bench 2.0, all 89 tasks, run with the Harbor harness, version 0.23.0.
Mel v0.1.62, run headless, with no person at the keyboard to answer it.
Auto Balanced, on Mel's live service.
Each task kept its own time limits. The run set no timeout multiplier.
49 tasks ran on October 24 and passed 27. The other 40 ran on October 25 and passed 19.
Not passed
A second run
We then ran the 43 tasks that Auto Balanced did not pass once more, on Auto Genius, with one try each. Auto Genius passed 21 of those 43.
This is a separate run of a different mode on part of the set. It is not a score for Auto Genius on all 89 tasks, and we do not add it to the Balanced score.
How to read it
The official Terminal-Bench 2.0 leaderboard is not taking new entries now, and it asks for at least five tries per task. Our run used one try per task, so it is not a leaderboard score.
We state the version, the mode and the setup so you can read the number for what it is.
Get started
The terminal is free to download. The agent is free on your own provider key.