Early access

Latch · San Francisco

The era of free tokens starts here.

We're building the agent machine: new silicon and software that run AI agents by the thousands, all day, without a meter running.

73,000tok/sMax throughput
~18WPower
1,000+Concurrent agents
5 × NVIDIA H100
1 Latch chip
Same tokens per second~73,000 tok/s
For the price of ~12 AirTags
The size of a coin~20 × 21 mm die

01 Mission

Stop counting tokens.

I opened an AI agent on a free plan and asked what it could do. That one question used 10% of my monthly quota.

What would you build if tokens were free?

Unlimited

No cap. Forever.

Your assistant runs all day, every day. No plan, no quota, no meter ticking.

Scale

Thousands of agents, one job each.

One per customer, one per repo, one per inbox. Specialists, not one tired chat.

Robotics

Robots that think in real time.

Today's models are smart enough. They're too slow and power-hungry for a body. Until now.

Protocol

Agents that talk to agents.

Skip email. Your agent asks mine, and only gets the data and people you approved.

Privacy

Your keys stay home.

No pasting passwords into someone else's cloud. Agents run on hardware you own.

Everywhere

AI where it never paid off.

Every camera, every machine, every classroom. Places a token bill never made sense.

02 Agents at scale

A thousand agents need a new operating layer.

Nobody can watch a thousand agents one by one. Today's tools were built for one chat at a time. When agents are as cheap as programs, they need what programs got decades ago: a harness that runs them, a protocol to talk to each other, and an interface that hands people only the decisions that matter.

Agent network · illustration0 agentsHover: you are the human

Imagine

Every customer with a problem gets their own agent on a call.

  1. Step 1Joins the call

    Hears the complaint, live, by voice.

  2. Step 2Debugs it

    Reads your logs and real data, not a summary.

  3. Step 3Fixes and replies

    Answers the customer and files the fix.

  4. Step 4Hands you the call

    Only when it needs a human decision.

Runs on your machineNo shared credentialsYour data never leaves$0 per call

03 Guarantee

No nerfs. Ever.

The model you get on launch day is the model you keep. It's in the silicon.

  • Quietly downgraded before the next release
  • Rate-limited when everyone logs on
  • Swapped for a cheaper model behind the same name
  • Repriced overnight

04 Superintelligence

Superintelligence isn't a higher score on a benchmark.

It's one person running a thousand agents at once.

Intelligence that multiplies what every person gets done. That only happens when running an agent costs nothing.

Early access, by invitation.

We're opening Latch to a small group of design partners first. If you got here from an email from us, that email has everything we're sharing right now, including any materials attached to it. Questions? Reply to it; it reaches the founder directly.

Want early access, or think you'd be a design partner? Write to us.