How are you using agent.terminal_init_command with Terminal Threads?
I’ve been trying Terminal Threads in Zed for CLI-based coding agents, and the setting that clicked for me is agent.terminal_init_command.
Instead of hardcoding one agent, I’m using it as an opt-in launcher in my dotfiles-backed Zed config:
{
"agent": {
"terminal_init_command": "agent=$(printf 'claude\\ncursor-agent' | fzf --prompt='Agent: ' --height=4 --reverse --no-info) && exec $agent"
}
}
When I open a Terminal Thread, I get a small picker and choose whether I want claude or cursor-agent for that session.
I like this because Zed still owns the workspace/thread surface, but the selected CLI owns the agent behavior: auth, config, slash commands, hooks, MCP setup, notifications, etc.
I still think ACP external agents are useful when I want Zed-native session management and tighter editor integration. But for some workflows, I prefer Terminal Threads because the CLI stays exactly itself. That means fewer surprises around CLI-specific behavior like slash commands, hooks, MCP config, auth, and notifications.
I’m also thinking this could go beyond a picker. The init command could inspect $PWD, choose a default based on the workspace/repo, and fall back to a picker elsewhere.
Curious how others are using Terminal Threads:
- Do you launch one default agent automatically?
- Do you use a picker?
- Are you doing directory-aware initialization?
- Are you mixing ACP external agents and Terminal Threads for different workflows?