Neuro Agent
Getting Started
Automation
Reference
Neuro Agent
Give it a goal. The agent plans, runs real commands on your device, reads the output, and keeps going until the task is done — with safety guardrails.
What It Does
Neuro Agent is a goal-driven autonomous shell agent built into your terminal session. You describe what you want to know or do, and the agent runs a loop:
At each step the agent picks a diagnostic command, explains why it's running it, executes it on the device over your existing transport (Serial, SSH, or WSL), reads the response, and decides whether it needs another step or has reached an answer. When it's done you get a concise summary.
Safety Model
The agent runs real commands on real hardware, so safety is the first-class concern. Every command the agent proposes falls into one of three buckets:
Safe · auto-run
Read-only commands on the built-in allowlist (ls, cat, ps, dmesg, ss, ip, journalctl, plus Zephyr shell and U-Boot read commands) run automatically without prompting.
Unknown · ask first
Anything the agent wants to run that isn't on the allowlist triggers an in-panel confirmation dialog. You see the exact command and the agent's reasoning before it executes.
Dangerous · always blocked
Destructive commands (rm, dd, mkfs, reboot, flash erase, anything with shell chaining like | or ;) are on a blocklist that always requires explicit user approval — even if you disable confirmations for other commands.
A hard step limit (default 10) prevents runaway sessions, and a max command length protects against serial buffer overflow on target devices.
How to Access
Two ways to start an Agent session. Both share the same step log and history.
@# in the terminal
Press @ to open Magic Input, then type # as the very first character inside the prompt to switch from AI chat to Agent mode. Type your goal and hit Enter.
Agent tab in the bottom panel
Open the bottom panel and select the Agent tab. You get a goal input, a step log, Play/Stop controls, and a history of prior sessions for the current terminal.
Example Goals
Neuro Agent shines at questions that need two or three commands chained together — the kind of thing you'd normally do by hand.
- Find what is listening on port 8080 and confirm which process it is
- Check whether the MQTT broker is reachable and which client IDs are connected
- List USB devices and tell me which driver is bound to each
- Why did the last boot take longer than usual?
- Show the current IP configuration and active routes
- Find the largest files under /var/log
Follow-ups
When the agent finishes a goal, you can ask a follow-up without starting over. The original goal and every prior step stay in context, so the next question can build on what the agent already found. Each terminal session keeps its own agent history.
Local or Cloud
The agent works with either backend. By default it uses your local LLM, so sessions stay fully offline. If you enable Cloud AI (OpenRouter), you can route agent reasoning to a larger model for harder multi-step tasks. The backend choice applies to every step of the agent loop.