About Perch

We build agents, and we study the models behind them.

Perch is a small team building an AI agent you can hand real work to. Alongside it we run research: we serve open models on our own GPUs with real agent work, measure what holds up, and publish what we find.

  • Web
  • Desktop
  • Terminal
  • Cloud
  • Phone
Perch running on a Mac desktop, in a terminal, in the browser and on a phone at the same time
Real app · one thread on every surface

Agents

Hand it the job, not a prompt

Research

Models and hardware, tested on real work

Open models

Run on our own GPUs to learn

Published

What real traffic teaches us

How we got here

An agent company that does its homework.

  1. Nov 6 to 8First free, unlimited weekend, served on our own B300, MI355X and H200 GPUs.
  2. Nov 9Our first serving study: what a weekend on our own GPUs taught us.
  3. Nov 11The Unlimited pass launches.
The announcement card for the Unlimited DeepSeek weekend
Nov 6 to 8 · the first unlimited weekend

What we build

Hand it a job. It runs a team.

Perch splits big work across helper agents, runs them at once and checks what they found before it answers.

  • Flock
  • Sub-agents
  • Routines
  • Cloud agents
  • Remote Control
  • Local models
Real recording · subagents launch film · 35 s

What we believe

Four rules we build by.

01

Agents should finish the job.

Not suggest, not draft. Do the work, check it and hand it back done.

02

Learn by running the models.

We serve open models on our own GPUs to see what really works for agents, then share it.

  • Open models
  • Our GPUs
  • Published

03

Real traffic is the benchmark.

We test models and hardware on the work people give Perch, and publish the numbers.

04

Your work stays on your machine.

Desktop and the CLI work on your own folders, and Perch can run on a model on your own computer.

Perch desktop app working through a local folder with three helpers running

Who we work with

Studies from production, with the people who build the hardware.

We run joint studies with GPU clouds and chip makers on real agent traffic, not synthetic benchmarks. If you run hardware or serve models and want real-world numbers, talk to us.

Talk to us