Terminal

Mobile Agent integrates with Termux, the Android terminal emulator, so the agent can run real shell commands and stream the output back into the chat in real time.

How it works

• The app sends a command over a local TCP connection (host / port / token) to the Termux streaming plugin.
• Termux executes it using its normal process environment, including any packages you've installed (pkg install, Python, Node, git, etc.).
• Output is streamed back as events — you watch the agent work line by line instead of waiting for a single blob.

Setup

1. Install Termux (from F-Droid or GitHub).
2. Install the Mobile Agent streaming add-on that exposes the TCP bridge (termux-stream).
3. In Mobile Agent, open Settings → Terminal and enter the host, port, and token shown by the bridge.
4. Tap Test — a connection change event confirms the bridge.

Once connected, terminal tools become available in the agent's toolset (subject to approval).

Events and UI

The terminal stream emits typed events:

• output — appended command output.
• done — command finished (with exit status).
• error — command failed.
• connectionChange — bridge connected or disconnected.

The Terminal tab renders a live, Ghostty-powered terminal view (expo-libghostty UI). Each executed command is also captured and displayed as a read-only command card inside the conversation (with its command, output, and taskId), so the chat history stays understandable after the fact.

What the agent can do through it

Because it's a real shell, the agent can:

• Install packages, run scripts, and build artifacts.
• Run git commands against a repository in the workspace.
• Interact with CLIs (the same ones you'd use by hand).
• Long-running work is streamed, and the scheduler can run terminal-based tasks in the background.

Safety

Terminal access is the most powerful capability in Mobile Agent, and it's treated that way:

• The tool approval flow gates every command — approve or deny per command, or enable autoApprove only for conversations you trust.
• Permissions are per-tool and per-agent (Tool Approvals, Custom Agents).
• The connection requires a token you configure — random, unguessable, and stored on-device.

Recommendation: keep terminal tools on ask unless you explicitly trust the conversation's agent and prompt.