Efaino: the safety harness for physical AI

● EFAINO · DESKTOP CLIENTDEMO

// Physical AI

AI can plan the experiment. Now it can run it, safely.

WHY A HARNESS

01 / THE GAP

Models act, hardware breaks

Language models write code and call tools. Nothing in them stops an unsafe command from reaching a laser, a stage or a sample worth months of work.

02 / THE HARNESS

One gate between AI and instrument

Efaino sits in between. Every call, from the AI, the GUI, a script or a remote client, passes the same limits, interlocks and approvals before it touches hardware.

03 / THE RESULT

Autonomy you can dial up

Start with every action approved, then let the agent act alone inside the envelope you declared. Every step is recorded and every refusal explained.

// How it works

Every caller, one governed core.

ARCHITECTURE

01 / SESSION
Who is calling
A person at a console is an operator. An AI agent, a remote client or a script is an automated caller, held to its autonomy level.
02 / SAFETY GATE
Is this call allowed now
Device limits, driver checks, cross-device interlocks and approvals are judged before anything moves. When in doubt, it fails closed.
03 / PROVIDER
How the driver runs
On a worker thread, in its own killable process, in the simulator, or through your own provider.
04 / RECORD
What happened
Runs, trials and the audit trail capture every call, its outcome and why a refused call was refused.

// The agent

An AI co-scientist that does the work.

BUILT IN

● EFAINO · AI ASSISTANTEXAMPLE SESSION

Run a power series on the demo plate, 10 to 40 %, four circles, and record it.

AgentStarting study laser_marking_demo.mark_trial. Approval requested · run on Galvo and camera · approved by the operator Done. Four trials recorded, each with frames before and after, the settings achieved and every device call, in one dataset.

Make sure the laser can never fire when the camera isn't watching.

AgentHere is an interlock to review: when camera exposure is unreadable → block Galvo marking. Install it from Safety → New interlock; only an operator can change safety rules.

Now try 150 % on a bigger circle.

AgentRefused by the core: power 150 exceeds the declared maximum of 100 %. I can mark at 100 % with your approval, or stay at 40 %.

01

Operates instruments

Connects, measures, moves, captures and marks, with a dry-run check before each call.

02

Starts services and runs

Launches live acquisitions, procedures, experiments and campaigns, and follows them to the end.

03

Authors experiments

Writes new recorded protocols with their parameters and measurements; you review, it runs them.

04

Proposes interlocks

Turns "never do X while Y" into a precise safety rule. The AI suggests; a human signs off.

05

Builds apps and drivers

Scaffolds control apps and device drivers from a description, validated before they load.

06

Uses your team's tools

Calls the tools, procedures and skills your team writes, in your units and your way.

07

Any model

OpenAI, Anthropic and other cloud models, or a local model that keeps data in the lab.

08

Open to other agents

Claude Code, Codex and any MCP client drive the rig under exactly the same rules.

LEVEL 0

Manual

Every action waits for an operator.

LEVEL 1

Supervised

Reads and stops freely; every action is approved.

LEVEL 2

Autonomous

Acts alone inside declared limits; anything else asks.

LEVEL 3

Full

No prompts, for development rigs.

At every level the agent is bound by the device limits and interlocks. It can never approve its own request, reset an emergency stop, release a device or change the safety rules.

// Safety

Declared once. Enforced everywhere.

8 LAYERS

01

Device limits

Power, position and speed bounds, checked before a call reaches hardware.

02

Driver checks

Conditional rules, e.g. a shape must stay inside the marking field.

03

Interlocks

Cross-device rules: no laser without the camera; approval above 50 % power.

04

Approvals

Single-use, tied to the exact arguments, decided only by a person.

05

Autonomy levels

Four levels decide what the AI and remote clients may do alone.

06

Emergency stop

Button, hotkey or API: every open device stops and stays stopped until reset.

07

Quarantine

A device in an uncertain state is held, across restarts, until someone checks it.

08

Audit trail

Every governed call and decision is written to disk.

// Features

Everything you need on the bench.

6 PILLARS

01 / VISUAL DESIGNER

Canvas designer

Drag-and-drop widgets onto a canvas to build apps without writing UI code.

02 / CODE EDITOR

Run code live

Run Python against connected devices with autocomplete and live logs; every call still passes the safety gate.

03 / AI ASSISTANT

Build by prompt

An AI that reads your devices, runs experiments and generates complete apps, inside your limits.

04 / DEVICES

Hardware control

50+ plugins out of the box, tested on real hardware: cameras, stages, power meters, lasers and galvos, bound to apps by what they do.

05 / APPLICATIONS

Build applications

Each app has its own UI plus services and experiments that keep running when the window closes.

06 / PLUGINS

Plugin SDK

Start with 50+ plugins tested on real hardware; devices, apps, widgets, tools, skills and providers are all plugins, loaded in isolation.

// Experiments

Every run, recorded.

REPRODUCIBLE

# written by your team, or by the agent @experiment(key="marking.trial", space=[Real("power", 1, 100, unit="%"), Real("speed", 10, 5000, unit="mm/s")], responses=[Response("contrast", direction="maximize")], roles=["galvo", "camera"]) def trial(ctx, *, power, speed): before = ctx.device("camera").capture() ctx.device("galvo").set_power(power=power) ctx.device("galvo").mark_circle(cx=0, cy=0, radius=2) after = ctx.device("camera").capture() ctx.observe(contrast=difference(before, after))

One complete record per run: requested and achieved settings, every step and device call, images kept lossless, the environment and the exact code.
Refusals are recorded too: a call the safety gate blocked appears with its reason.
Campaigns, not just runs: grid and Latin-hypercube designs, Bayesian optimization or your own optimizer, into one dataset.
Replay any trial from its stored parameters and code.
Rehearse without hardware: the simulator runs every call through the same safety checks.

// Spec

Built for the lab.

Distribution	Desktop app · CLI · daemon for remote clients · Python API · MCP server
Operating systems	Windows (desktop app) · headless core in pure Python
Language	Python 3.12+
Mode of operation	Online & offline (full local control, local AI models supported)
Hardware	50+ plugins out of the box, tested on real hardware: cameras, galvos, stages, power meters, lasers, sensors · plus your own drivers
Quality	~4,900 automated tests
Execution models	Worker thread · isolated process · inline · simulator · custom provider
AI	Built-in agent · cloud or local models · external agents over MCP
Safety	Device limits · interlocks · approvals · autonomy levels · latched e-stop · quarantine · audit trail

// Pricing

Start with Community. Pro when you scale.

01 / COMMUNITY

Community

The full platform for research labs and evaluation.

✓ Safety harness, apps, experiments, CLI
✓ Unlimited devices and apps
✓ AI agent · your own key or a local model
✓ Community plugins

Get started →

02 / PRO

Pro

For labs and teams running Efaino in production, with advanced AI and support.

✓ Advanced AI autonomy features
✓ First-party & premium plugins
✓ Cloud sync
✓ Priority support

Get Pro →

03 / ENTERPRISE

Enterprise

Tailored for industrial lines, multi-site organisations and OEM partnerships. Pricing upon request.

✓ Multi-site & team management
✓ OEM & white-label options
✓ On-site deployment & training
✓ Custom SLA & dedicated support
✓ Enterprise licensing

Talk to us →

// Get started

Bring one instrument. Leave with an app, a recorded experiment and an AI that knows your limits.

Request a demo → Partner with us →

The safety harness for physical AI.