agenterm

Run dev servers, watchers, builds, and tests in terminals that you and your coding agents can watch, type into, and stop. Each command gets an id, so an agent never starts the same server twice.

npm install -g agenterm

agenterm run --id web "npm run dev"    # start it (or get the one already running)
agenterm ls                            # what is running, with ports
agenterm logs web -f                   # follow its output
agenterm stop web                      # Ctrl+C, then end it

Or without installing: npx agenterm ls.

Tasks keep running in a small background daemon after the terminal or agent that started them exits. The agenterm desktop app shows every task as a live terminal. On a Mac it also includes this CLI and adds the agenterm command for you, so you do not need this package as well.

For agents (MCP)

claude mcp add --scope user agenterm -- agenterm mcp   # Claude Code
codex mcp add agenterm -- agenterm mcp                 # Codex

Other agents (Cursor, Claude Desktop, …) take this in their MCP config:

{ "mcpServers": { "agenterm": { "command": "agenterm", "args": ["mcp"] } } }

Without a global install, use "command": "npx", "args": ["-y", "agenterm", "mcp"].

The tools match the CLI: run, list, logs, input, stop, remove, processes, and kill.

Commands

| Command | What it does | | ------------------------------------ | --------------------------------------------------------------------- | | run [--id ID] "<command>" | Start a command in a new terminal, or return the one already running | | ls [--running] [--json] | List tasks, running first | | show <task> | One task in detail | | logs <task> [-n N] [-f] | Recent output; -f keeps printing until it ends | | input <task> <text> | Type into a task, then press Enter (unless --no-enter) | | stop <task...> [--force] | Ctrl+C, then end the process tree; --force kills right away | | rm <task...> | Remove tasks and their output, stopping them first | | clear | Remove every task that is no longer running | | ps [--orphaned] [--all] [--json] | Commands running outside agenterm: orphaned, in terminals, under agents | | ps --hide <name\|pid>, --unhide | Leave a program out of ps (it keeps running); --all shows it | | kill <pid...> [--force] | End a command from ps and everything under it | | settings | Show or change settings, e.g. --stop-with-agent on | | daemon start\|stop\|restart\|status | Manage the background daemon | | mcp | Serve the same tools over the Model Context Protocol (stdio) |

Updating

npm install -g agenterm, then agenterm daemon restart if tasks are running. Once a day, in a terminal, the CLI asks npm whether a newer version is out and says so after the command. It never does for mcp, --json, or CI; AGENTERM_NO_UPDATE_CHECK=1 turns it off.

Requirements

• macOS or Linux (x64 or arm64; on Linux, glibc rather than musl, so not Alpine), Node.js 22 or newer. Everything installs prebuilt: nothing compiles, and npm runs no install scripts.
• Installed with a different version than the daemon already running? agenterm daemon restart swaps it (the CLI does this by itself when nothing is running).

Data lives in ~/.agenterm/. The daemon listens on 127.0.0.1 only and every request needs a token stored there.

MIT licensed. Website: https://agenterm.entange.com. Source, issues, and the desktop app: https://github.com/entange/agenterm.