Embroid
Solution

Give AI agents safe access to real hardware.

Give AI coding tools safe, scoped access to hardware feedback through MCP and APIs.

The problem

AI coding tools can generate firmware, but they need safe, structured feedback from actual devices — and permissions that keep them from bricking a bench.

An AI coding agent can iterate fast on code, but without real-device feedback it's guessing about behavior. The naive fix — giving an agent shell access to hardware tools — is how boards get bricked and labs get wedged.

How Embroid helps

Embroid exposes controlled MCP, CLI, and API tools so agents can flash, observe, reset, and iterate with guardrails that keep humans in the loop.

Agents talk to hardware through a typed MCP surface. Every verb declares its permissions. Leases are short and explicit. Humans approve destructive scopes. The agent gets the feedback it needs; the lab keeps its guardrails.

The loop

Request. Act. Observe. Prove.

Agents and humans use the same four verbs on real devices. Embroid makes each one safe, scoped, and recorded.

01
Request

Reserve a board from the lab

The agent requests a scoped lease on a specific target, with the verbs it's allowed to call and how long it has.

02
Act

Flash, run tests, power-cycle

Within the lease window, the agent flashes firmware, runs sequences, toggles power, and triggers accessories.

03
Observe

Capture serial, video, sensor data

Every command produces structured output — logs, traces, captures — streamed back into the session record.

04
Prove

Verify behavior with captured evidence

The session record is signed and exportable. Reviewers replay or audit exactly what the agent did and why it passed.

What it looks like

Agent fixes a flaky Bluetooth pairing bug.

A coding agent picks up a flaky Bluetooth pairing issue from the backlog, reserves a board, ships a candidate fix, and proves it works — without a human at the bench.

  1. 01Agent reserves the target board and takes a 15-minute lease.
  2. 02Pushes the candidate firmware and flashes the device.
  3. 03Runs the pairing sequence; streams serial output back into the agent's context.
  4. 04Captures the pass/fail signal and attaches the signed session record to the PR.
  5. 05Lease expires; the board returns to the queue, ready for the next run.
Outcomes

What teams get from Agent-Assisted Firmware

  • Safe, typed hardware tools through MCP
  • Scoped permissions for destructive actions
  • Structured failure feedback for fast iteration
  • Audit logs that show exactly what the agent did
Products used

Which Embroid product fits this workflow

Embroid Client

Great starting point for individual engineers working with coding agents at their own bench.

Explore Client
Embroid Basic

Keeps the agent's target bench online overnight for long iteration loops.

Explore Basic
Embroid Pro

For teams coordinating multiple agents across a shared lab.

Explore Pro
Platform capabilities

Built on these parts of the platform

Example evidence

What the record looks like

Every agent-driven session is fully replayable: the capability issued, the calls made, the logs captured, and any human review events. Audit trail is first-class, not bolted on.

Implementation path

A realistic rollout

Week 1

Install Client and wire MCP into the engineer's coding agent.

Week 2

Define agent capabilities and run the first agent-driven flash-and-observe loops.

Week 3+

Expand to team-wide agent workflows with Basic or Pro as the bench becomes central.

Agent-Assisted Firmware

Make this workflow real on your hardware.

Share your setup. We'll help you pick the right Embroid product and get you to a first working run.