AI agents

Let AI agents work on real hardware

AI coding agents are good at writing code, but embedded code only proves itself on the board it runs on. NablaBox gives an agent the same access an engineer has – power, serial console, test runs and logs – through its REST API, and the same limits: its own account, only the DUTs you assign, every action logged.

NablaBox includes an MCP server, nablabox-mcp. Your agent’s MCP client starts it like any other tool server, and the agent can use the board right away – no glue code, no prompt engineering around a REST API.

What the agent can do

Fifteen tools cover the full work on a board. DUTs are named by name or slot number.

Power

Switch a DUT on or off, or power-cycle it – a hard reset with nobody at the bench.

Serial console

Open the console, type commands and wait for a prompt such as login:. Output arrives as clean text, without terminal escape codes.

Test runs

Run tbot test cases with parameters, wait for the verdict and get the run id and the end of the output. A run that hangs is stopped at its time limit.

Logs and history

List past runs with their results, read the full log of any run and the DUT’s own event log.

Test catalogue

Discover the test cases on the device – the shipped suite and your own – with their parameters and defaults.

DUT status

See every DUT the account can read: power state, voltage, console settings and whether a test is running.

The loop

The agent works in the same cycle as an engineer at the bench – only without waiting for someone to walk over and press reset.

1. Write

The agent writes or changes a tbot test case for the DUT.

2. Run

It power-cycles the DUT and runs the test against the real board.

3. Read

It reads the full run log, the verdict and the console output.

4. Fix

It fixes the test – or finds the bug in your software – and runs it again.

Limits enforced by the device

An agent’s instructions can be misread; the device’s permissions cannot. The account’s limits are enforced by NablaBox itself, whatever the agent sends – and the MCP server adds its own on top.

Assigned DUTs only

The agent’s account can only power, console and test the DUTs you assign to it. Requests for any other DUT are rejected.

No admin rights

With the user role the agent cannot reboot or reset the gateway, change the network, or manage users.

DUT tools only

The MCP server offers no tools for users, GPIOs, networking or system settings at all – not even with an admin account.

Passwords stay out

A DUT’s login password is withheld from the agent unless you choose to share it, so it doesn’t end up in transcripts.

No runaway runs

A test run that exceeds its time limit is stopped and the DUT is powered off – no board left running all night.

Full audit log

Every request is logged with user, action and result – including the rejected ones. You can see exactly what the agent did.

Isolated DUT network

DUTs sit on their own network segment, never bridged to your company network unless you enable it.

One console session

Each DUT’s console admits one session at a time, so an agent and an engineer never type into the same console.

No inbound ports

Remote access runs over an outbound NetBird connection. The agent can work from anywhere, your lab firewall stays closed.

Connect your agent

nablabox-mcp runs on the agent’s machine – Linux, macOS or Windows – and connects over the NablaBox’s own HTTPS, the same address as the web UI. The unit’s certificate is checked by its fingerprint, like an SSH host key, so nothing else can pose as your NablaBox. For Claude Code:

nablabox-mcp --host your-nablabox \
  --show-fingerprint

claude mcp add nablabox \
  --env HWDAMON_PASSWORD=... -- \
  nablabox-mcp --host your-nablabox \
  --tls-fingerprint AB:CD:... \
  --user agent

Then just ask: “Power-cycle dut0, wait for the login prompt and run the temperature test.” Any MCP client that starts local servers works the same way.

Case study: our own test suite

The generic test suite that ships on every NablaBox – 49 tbot test cases for U-Boot and Linux – was written by an AI coding agent (Claude Code), against real boards on a NablaBox. It covers boot, memory, storage, network, watchdog, secure boot, SWUpdate and SoC register checks, with register maps for the i.MX8MM, i.MX8MP and i.MX95.

The real hardware made the difference. On an i.MX8MM, kernel messages on the serial console landed in the middle of register reads, and a dump of 381 registers failed partway. The agent’s fix – quieting the console during reads and retrying garbled lines – was tested on the same board. No simulator would have shown that bug.

Or use the REST API directly

Log in as the agent’s account, power-cycle a DUT and read its test results:

NB=https://your-nablabox
TOKEN=$(curl -s -X POST $NB/api/login -H 'Content-Type: application/json' \
  -d '{"username":"agent","password":"..."}' | jq -r .token)
AUTH="Authorization: Bearer $TOKEN"

curl -s -H "$AUTH" $NB/api/duts                        # list the DUTs
curl -s -X POST -H "$AUTH" $NB/api/duts/dut0/power/off
curl -s -X POST -H "$AUTH" $NB/api/duts/dut0/power/on
curl -s -H "$AUTH" $NB/api/duts/dut0/tests              # test runs and verdicts
curl -s -H "$AUTH" $NB/api/duts/dut0/tests/RUN/log      # full log of one run

Consoles and live test runs use WebSockets on the same API. Give your agent these calls and it can do the rest.

Want to try it with your own boards?