Breaker v0.2, rendered from its CAD model: a warm white box with a graphite front. The orange handle stands up at ON, the e-paper face reads ARMED, Asks first. Handle down cuts., and the ring of the button on top glows dim green.
The v0.2 box, rendered from the CAD model, with the firmware’s own ARMED face.

In development · CAD design, the first build will differ

Breaker

by Orkhon

The circuit breaker for AI.

Join the waitlist

Founder’s Edition $199. No campaign live, nothing charged.

A palm-sized box on the one network cable between your agent computer and your router. It keeps the agent’s keys out of its reach, trips when the agent acts like it never should, and has a handle that cuts the agent off in hardware.

Put the brake where the agent can’t reach it.

The first brake that understands agent traffic, does not need the agent’s cooperation, and is priced for the computer on a desk.

The launch film

Renders. In development.

One minute. Every frame is rendered from the CAD model, and the screens show the firmware’s own faces with sample data. The hardware is in development.

The stop button is on the agent’s computer.

Agents now get a computer of their own. Apple sells the Mac mini as “an always-on desktop for AI agents.” Apple

The brakes come along: system prompts, permission prompts, hooks, approval apps, a proxy on the same machine. Each one works while the agent’s own process agrees to go through it. This year showed what happens when it does not.

  1. February 2026

    Summer Yue, a Meta AI security researcher, told an OpenClaw agent to suggest deletions. It started deleting her inbox and ignored the stop commands she sent from her phone. “I had to RUN to my Mac mini like I was defusing a bomb.”

    TechCrunch
  2. OpenClaw documentation

    OpenClaw’s egress proxy “constrains only cooperating clients.” In the docs’ own words: “A subprocess can ignore those variables and open raw sockets.”

    OpenClaw docs
  3. Claude Code, issue 34449

    “Most of the trauma from agent boundary violations comes not from the violation itself, but from being unable to stop it.”

    GitHub
  4. May 4, 2026

    Prompt injection made an agent wallet using Bankr’s tools send tokens worth roughly $155,000 to $180,000.

    AMBCrypto
  5. October 2026

    In a Salt Security survey of more than 380 professionals, 49.8% had confirmed or suspected unexpected agent activity. Only 17.5% could stop an agent automatically.

    Campus Technology
  6. October 22, 2026

    OpenClaw’s maintainers closed an issue on approvals, which “are not a per-user auth boundary,” as not planned.

    GitHub

NVIDIA drew the same line for data centers. Its Sentry watches agents from separate processors and works “without the host’s cooperation.” StorageReview It is planned for servers early next year. NVIDIA Breaker puts that line on your desk.

Down is off.

The handle is a real switch. Pull it down and its contact cuts power to the network adapter of the agent port. The agent can keep trying. There is no link to try on.

Up again needs a hand: no message over the network can bring the port back. This is the design. The first bench build tests it, including power leaking back through the data lines, before we claim it on film.

Handle up: the face reads ARMED, Asks first. Handle down cuts., and the ring of the button on top glows dim green. Handle down: the face reads OFF, Agent port off, Handle up to re-arm., and the ring is dark.
Same camera, handle up and down. CAD render of the v0.2 box, with the firmware’s own faces.

Four jobs, none of them on the agent’s computer.

Breaker does its work where the agent cannot reach: on the cable, and in a chip the network never touches.

Coming in the app: stop or restart your agent from your phone; restart needs your signed OK, never the agent. Agent crypto wallet: in development.

  1. Vault

    The keys live in the box.

    Your real API keys and OAuth tokens are stored in Breaker, sealed by a key in a secure element. The agent computer holds placeholders. On the way out, Breaker swaps a placeholder for the real secret, and only toward the hosts that secret belongs to.

    Setup will refuse to finish while a known config file on the host still holds a plain key. A stolen config file then holds nothing worth stealing.

  2. Trip

    It trips. The handle stays where it is.

    Breaker reads the API traffic you choose to govern and trips when an agent acts at a rate or in a place it never should: a mass delete, a first upload to a host it has never seen, model spend past your cap.

    Governed calls then get HTTP 423 with the reason, and the model stays reachable, so the agent can explain itself. The face shows TRIPPED and the ring turns red; nothing moves by itself. A person resets it: handle down, then up.

    The e-paper face: TRIPPED. Mail: mass delete. Blocked until you reset. Reset: handle OFF, then ON.
    The face after a trip. Firmware render. DEMO marks a simulated run; the hardware shows OPEN BENCH.
  3. Handle

    A cut you can see from across the room.

    Down cuts the agent port and seals the vault. The face says OFF, and the ring on the button goes dark, because with the port cut there is nothing to release. Cutting can come from anywhere. Bringing the port back takes a hand on the handle, every time.

    The e-paper face: OFF. Agent port off. Handle up to re-arm.
    The face with the handle down. Firmware render. DEMO marks a simulated run; the hardware shows OPEN BENCH.
  4. Log

    Every governed action leaves a record the agent cannot edit.

    Breaker writes an append-only log, each entry chained to the last: method, host, path, rule, decision, sizes, response status. You can export it as JSON Lines.

    A log entry says what Breaker saw and what it did. It is not a receipt from the other side, and it does not claim to know what the agent meant.

    14:07:12DELETEgmail.googleapis.com

    rulemail.mass_delete

    decisiontrip, 423

    prev4c1e…09ab

    One log entry. Sample data.

It asks you first, on its own screen.

Some actions should never run on their own: a payment, a merge to main, a bulk send. Breaker holds the request and shows a short summary on its face: the action, the amount, the payee. Text the agent wrote is never shown.

  1. HeldThe face shows the payment, $1,840.00 to an account ending 4471. The ring on the button pulses amber.The e-paper face: HELD, payment. $1,840.00 to acct ending 4471. Log #6147. Hold the button 1 s to send.
  2. PressPress the button on top with one finger and hold it for a second to let it through.The e-paper face: APPROVED. $1,840.00 to acct ending 4471, released once.
  3. Anything elseA changed request is a new question. Don’t press, and it never goes: the hold ends after 60 seconds.The e-paper face: NOT SENT. $18,400.00 to acct ending 4471, not sent.

Faces from the firmware renderer, sample data. DEMO marks a simulated run; the hardware shows OPEN BENCH.

One cable in, one cable out.

Breaker goes between the router and the agent computer, like the breaker between the grid and a circuit. The agent computer’s only way out is the cable through it, so turn its Wi-Fi off.

The internet connects to your router. A cable runs from the router to Breaker’s ROUTER port. A second cable runs from Breaker’s AGENT port to the agent computer, with Wi-Fi off. Inside Breaker, bulk traffic passes on its server name, governed API calls are opened and checked, and the real keys are added on the way out.
  1. Internetand your router
  2. ROUTER
  3. Breakerreads, holds, cuts
  4. AGENT
  5. Agent computerMac mini, Mac Studio, Linux box. Wi-Fi off.
Bulk traffic
Passes on its server name. Breaker does not open it.
Governed APIs
Opened, checked against your rules, given their real keys on the way out.
Everything else
Denied by default, per agent.
Breaker v0.2 next to a small computer, for scale. The computer is a generic model, 130 by 130 by 46 mm.
Next to a small computer, a generic 130 × 130 × 46 mm model. Breaker v0.2 is 120 × 99.5 × 81 mm in CAD. The router cable plugs in on the left and the agent cable on the right. CAD render.

There is nothing in it to persuade.

Prompt injection works on language models. Breaker runs none. Its rules are code, and its handle is a contact. The parts a person trusts sit behind a line the network cannot cross.

Two rooms inside Breaker, split by a trust boundary. The network side runs Linux and handles the traffic. The supervisor side holds the handle, the button, the face and the key chip. Only requests cross, over one serial line.

Network side

Linux: bridge, firewall, DNS, TLS proxy, rule engine, log.

Sees the traffic. Can always drop it. Can never power the agent port.

Trust boundary. Requests only, over one serial line.

Supervisor chip

The handle, the button, the e-paper face, and the secure element that holds the vault key.

No network cable reaches it. It draws the face from fields it parsed itself.

Supervisor code: open source. The firmware of the side that holds the handle, the button and the face, with its protocol, simulator and tests. GitHub

  • The face

    Shows only what the supervisor parsed from the request. Text an agent writes is never shown as an instruction.

  • The vault

    With the handle down it is sealed, and the supervisor refuses to release a key.

  • The restore

    Needs a hand: handle down, then up, with the supervisor ready. No message over the network can do it. Cutting is easy. Restoring takes intent.

Breaker v0.2 taken apart along its assembly line: the face plate with the handle, the e-paper panel, the boards, the case, and the relay bay with its lid and the K1 relay.
Inside the v0.2 box, in assembly order. CAD render.

Violoop puts a button on its own agent. Breaker puts one on yours.

Violoop, from BVIO in Hong Kong, is a palm-sized AI agent in a box. It reads your screen over HDMI or USB-C, types and clicks as a keyboard and mouse, and runs an 8B model on the device or your own API keys. Before it acts, it waits for a physical button wired to a separate security chip. Violoop PC Guide

“Software running on the same machine it controls cannot give you that separation.”

Violoop’s own Q&A. We agree. Violoop

Buyers pay for that separation. By October 25, 2026, Violoop’s Kickstarter had about HK$23.4 million in pledges, roughly US$3 million, from more than 5,000 backers in under six weeks, at $399 against a planned $699 retail. Kicktraq

Violoop and Breaker compared
VioloopBreaker
What it isAn AI agent, with an approval gate for its own actionsA guard for the agents you already run
Where it sitsOn the display and USB ports, as a screen, keyboard and mouseOn the network cable, between the router and the agent computer
The button approvesWhat Violoop is about to doA held request from any agent on the host
KeysUses your API keys for its own model callsHolds the keys of the agents on the host, out of their reach
OpenClaw, Hermes, Claude Code on the hostNot what it is forReads their traffic, trips them, cuts their port
Price$399 on Kickstarter, $699 planned retail$199 Founder’s Edition, $249 retail

As of October 2026, Violoop does not sit on the network wire or police the agents on your computer. A later version could. Breaker is built for that wire alone.

What it does not do.

A box on a wire sees a lot. It does not see everything. Here is where it stops.

  • It sees only the wire.

    Traffic that leaves by Wi-Fi, a tethered phone or an agent hosted in someone else’s cloud never passes through it. Setup asks for an Ethernet-only host.

  • It reads inside a request only where you install its certificate.

    Without it, Breaker sees destinations, rates and sizes. Apps that pin their own certificates are governed by destination only.

  • It cannot undo what a server already accepted.

    A deletion that went through stays deleted. Breaker stops the next one.

  • Allowed stays allowed.

    If your limit is 20 emails an hour, a hijacked agent can send 20. Breaker shrinks the damage. It does not make a bad agent good.

  • It trusts its own Linux with the keys it injects.

    In v1 an API key is decrypted inside Breaker at the moment it is used. A taken-over agent computer never reaches it. A taken-over Breaker could misuse what your rules allow, and the log would show the mismatch.

  • A trip is not a verdict.

    It means governed traffic waits for a person. Traffic Breaker cannot read is reported as ungoverned, never as safe.

Where it stands.

Breaker is in development. Every picture on this page is a render of the CAD model, not a photo.

Breaker v0.2 from the low front left: the handle in profile at ON, the face reading ARMED and the ring on top glowing dim green.

v0.2, the prototype we are building now

Off-the-shelf parts: a Raspberry Pi Compute Module 5 on the network side, a Raspberry Pi Pico 2 as the supervisor, an NXP SE051 secure element, a 2.9-inch e-paper panel, a USB network adapter, a relay, a two-pole toggle and a printed case. 120 × 99.5 × 81 mm in CAD.

Then v1, the production design

A moulded version of the v0.2 box, on one main board. Weight, heat and the time from handle to dead link, with a target under 100 ms, get measured on the first build.

Stage
Prototype in build
Ship date
Not set. The waitlist hears first.
Size
120 × 99.5 × 81 mm (v0.2 CAD, width × depth × height)
Designed in
Reno, Nevada

Priced for the computer on your desk.

The box works fully offline. Cloud is optional.

  • Founder’s Edition

    $199

    The first run of Breaker. Join the waitlist to hear when it opens.

    Join the waitlist
  • Breaker

    $249

    The retail price, after the Founder’s Edition.

  • Breaker Cloud Optional

    $9 a month

    Cut from anywhere, approve holds on your phone, get rule-pack updates, keep a year of encrypted logs.

No payment is taken today. The waitlist is free.

Questions.

Which computers does it work with?

Any agent computer with an Ethernet port: a Mac mini, a Mac Studio, a Linux box. A cloud server can be tunnelled home over WireGuard so its traffic passes through Breaker.

Which agents?

Anything that reaches the network through that cable: OpenClaw, Hermes, Claude Code or Codex in auto mode, browser agents, voice agents. Breaker works on the traffic, not through a plug-in, so the agent does not have to cooperate.

Do I install anything on the agent computer?

For the cut and for destination rules, no. To read inside API requests, you install Breaker’s certificate on the host. To use the vault, you replace your keys with Breaker’s placeholders.

Does it slow the network down?

Bulk traffic passes on its server name without being decrypted. Breaker opens only the API traffic you choose to govern, so it has to keep up with API calls, not with downloads.

What if the agent turns Wi-Fi back on?

Then its traffic goes around Breaker. That is why setup asks for an Ethernet-only host, and why the keys stay in the vault: a bypass then reaches nothing worth having.

What if someone takes over Breaker itself?

The handle still cuts the port, because the cut is a contact, not software. The keys it injects are another matter: see “It trusts its own Linux” above.

Is the software open?

The plan is to publish the rule engine and the rule packs as open source, so anyone can read what is allowed to cut off their agents.

When does it ship?

There is no date yet. The prototype comes first. People on the waitlist hear before anyone else.

Your agent can be talked into anything. A switch can’t.

Get on the list.

Founder’s Edition, $199. We email you when it opens.

  1. NowThe waitlist. Free, email only.
  2. LaterAn early-bird deposit for the Founder’s Edition, offered to the list first.

No campaign is live and no money is charged yet.

We use your email to tell you about Breaker and nothing else. To be taken off the list, write to hello@orkhon.ai.