Mel – the agentic terminal
The AI terminal for teams.
Learn · superintelligent terminal
What is a superintelligent terminal?
Updated
A superintelligent terminal is a terminal where AI agents do the work, not only answer questions. You describe a task, and a team of agents plans it, builds it, tests it and ships it, inside your real project. You stay in charge of what runs.
No terminal is superintelligent today. People use the phrase for the goal: a terminal you can hand bigger and bigger pieces of work to, and trust with the result. Mel – the agentic terminal is built toward that goal. This page explains what the phrase means, how Mel approaches it, and how to judge the claim.
What people mean by it
When people say "superintelligent terminal", they usually mean four things together:
- Agents, not autocomplete. The AI runs commands, edits files and reads the real output, not only suggests a line.
- A team, not one chat. One agent plans, others build and test their parts, and the work is checked before it reaches you.
- The best model for each step. It is not tied to one AI company.
- Proof, not adjectives. It shows results on named tasks, and it says what it could not do.
How Mel approaches it
Mel is a terminal for Mac, Windows and Linux with a team of AI agents built in.
- A team of AI agents in one terminal. A lead plans and assigns, and workers build and test their parts in parallel, each in its own pane. See What is a multi-agent terminal?
- Plans, builds, tests and ships. The agents run the build and the tests and read the real output before they call a task done.
- Many models. Claude, GPT, Gemini and more, or let Auto pick the model for each task. See What is a multi-model coding agent?
- Configurable approvals. You choose which commands and edits need your yes. Set approvals to Always ask and each change waits for you. A step that cannot be undone always asks. See What are AI agent approvals?
- Your terminal still works. Every command you type yourself runs as it always has.
Proof: a measured result
On Terminal-Bench 2.0, a public set of 89 hard terminal tasks, Mel passed 46 of 89 (52%). Setup: Mel v0.1.62, Auto Balanced, all 89 tasks, one try each, Oct 25, 2026. The score is self-reported.
A score is a measure, not a promise. It shows the agents can finish about half of these hard tasks with no help, and it shows how much is still left to do.
How to judge a "superintelligent terminal" claim
| Question to ask | A weak sign | A strong sign |
|---|---|---|
| Does it run the work, or only suggest it? | Suggestions you copy | Commands run, output read |
| Does it check its own work? | A summary of what it meant to do | Tests run, output shown |
| Can you control what it runs? | All or nothing | Approvals you configure |
| Is the claim measured? | Adjectives | A named benchmark with its setup |
| Does it say what it could not do? | Always reports success | Says what failed and why |
FAQ
What is a superintelligent terminal? A terminal where a team of AI agents plans, builds, tests and ships work for you, under approvals you set. No terminal is superintelligent today; the phrase names the goal.
Is Mel a superintelligent terminal? Mel – the agentic terminal is built toward that goal. Today it is a terminal with a team of AI agents that plans, builds, tests and ships, on many models, with configurable approvals.
How good are Mel's agents today? On Terminal-Bench 2.0, Mel passed 46 of 89 tasks (52%), with Mel v0.1.62, Auto Balanced, one try each, on Oct 25, 2026. The score is self-reported.
Which AI models does Mel use? Claude, GPT, Gemini and more. Auto picks a model for each task, or you can pin one.
Does every change wait for my approval? You decide. Set approvals to Always ask and each change waits for your yes, or let the agent run some steps on its own. Read-only commands can skip the card, and a step that cannot be undone always asks.
How do I try it? Download Mel for Mac, Windows or Linux. It is free with your own API key, and Pro and Max add hosted credits.
Mel – the agentic terminal · openmel.dev. A multi-agent, multi-model AI terminal for Mac, Windows, Linux and VS Code. Free with your own API key.