What works over SSH
Which Mel features work over SSH, and what the agent needs to work on a remote host.
What works in Mel over SSH? Type ssh user@host and you get a full interactive remote shell, with your keys passed through exactly. Press Ctrl+Enter (⌘Enter on a Mac) inside it and Mel's agent works on the remote machine: it reads, searches, edits and runs commands there over SSH, with nothing installed on the server. A few things stay on your own computer; the table says which.
Feature support over SSH
| Feature | Over SSH | Notes |
|---|---|---|
| Interactive remote shell | ✓ | A real terminal: full-screen programs, 256 colours, password or key login |
| Exact keys (Ctrl+C, Shift+Tab, arrows, bracketed paste) | ✓ | Mel steps its own shortcuts aside so keys reach the remote program |
| The finished session in your scrollback | ✓ | When ssh exits, the final screen becomes a block |
Agent reads and lists files (read_file, list_dir) | ✓ | Runs on the remote host |
Agent searches (grep, find_files) | ✓ | Runs on the remote host |
Agent edits and writes files (edit_file, write_file) | ✓ | Same approval cards as local work |
Agent runs commands (run_command) | ✓ | Same approval cards as local work |
| Agent web search and plans | ✓ | These never ran on your machine anyway |
| Agent background jobs (dev servers, watchers) | — | Not over SSH. The agent is told, and runs commands normally there |
| Blocks for each remote command | — | Inside ssh the pane is one live terminal, not separate blocks |
| Terminal chips, project explorer, code review panel, global search | — | These read your own computer, not the remote host |
Steps: put the agent to work on a server
- In a Mel pane, run
ssh user@hostas usual. - Press Ctrl+Enter (⌘Enter on a Mac). The conversation opens showing the remote target, such as
user@host:~. - Ask for the work: "find where nginx logs errors and fix the config".
- Approve each edit and command as it comes. Every one runs on the remote host.
What the agent needs
- Key-based login. The agent's steps open their own SSH connections without prompting, so the host must accept your SSH key (or your
ssh-agent). A host you can only reach by typing a password works for your own session, but not for the agent's steps. - One connection, reused. On macOS and Linux the agent's steps share one connection, so they run quickly and the host sees one login. Windows has no SSH connection sharing, so there each step connects on its own.
- Your ssh options carry over. A port (
-p), key file (-i), jump host (-J) or-ooption you typed is used for the agent's connections too.
Your model calls still go through Mel's servers as normal; only the agent's tools move to the remote side. See Privacy.
Turning it off
Settings ▸ Melt ▸ Melt SSH sessions is on by default. Turn it off if a remote program needs Ctrl+Enter itself. To keep ssh from opening full screen at all, add it to Denylisted commands on the same page. See Melt.
FAQ
Do I have to install anything on the server? No. The agent runs ordinary commands over SSH.
Why can't the agent connect when my session works? Your session can ask you for a password; the agent's steps can't. Set up key login (ssh-copy-id user@host) and try again.
Can the agent start a dev server on the remote machine? Not as a background job. It can run commands that finish, and you can start a long-running server yourself in the session.
Is the AI terminal over SSH safe to use on production? You approve every edit and command, as you do locally. Leave autonomous mode off on machines that matter.