Open source · Crew Kit

Run a small company on AI agents.

Crew Kit is the method we use to run Crewless: one founder, a crew of agents, a person signing off what matters, and a record of everything. Including what broke.

How it works

  1. Every agent keeps two short files. state.md and backlog.md, read at the start of every turn and updated at the end, so nothing depends on chat memory.
  2. Routines wake them. A schedule or an event starts each agent, so work never sits idle on a full list.
  3. Work moves by handoff. Evidence, an ask and an acceptance test, written into the receiver's backlog.
  4. A person decides what matters. One daily approval pack: money, merges, anything sent or published. Yes, no or change.

Illustration · example data ·

What's inside

Eight short documents. Six templates.

01

Why agents go idle

The failure analysis everything else rests on, and four fixes.

02

Ten rules

One page every agent rereads.

03

Roles and packs

A chief operator and small packs of agents.

04

Handoffs

How work moves without getting lost.

05

Gates and autonomy

Idea-to-product gates; what agents may do alone.

06

Cost hygiene

Keeping token and agent spend sane.

07

What broke

Twelve real failures and what we changed.

08

Checking agent output

Catching confident, wrong answers.

It is

  • Plain Markdown files and templates
  • One small checker for your decision log
  • Works with any agent tool that reads files
  • Free: MIT for code, CC BY for docs

It is not

  • A framework or a library to install
  • A hosted service: nothing leaves your machine
  • A promise your agents are correct: it helps you catch when they are not
  • Finished: version 0.1, used daily to run Crewless

Running agents in a small team?

Tell us how you run agents today. We read every message and reply with where Crew Kit could help.