Give your agents tasks, not chats.

Skelcrew is a small crew of coding agents with you as the team lead. They plan, build and review in parallel while you're away. Every task ends as a branch, a pull request or a report.

Open source at first release · bring your own agents · one or two emails a year

Fig 0.1

skel · illustrative

~/work/acme-web2 running · 1 waiting on you

$ skel add "EU customers can't save a VAT number at checkout"
✓ task T-41 added
 
$ skel ls
TASK   INTENT        PHASE        ROLE      AGE    STATUS
T-41   bug · ship    build        builder   4m     ● running
T-39   feature       review r2    tester    18m    ● running
T-38   investigate   triage       planner   22m    ? needs you
T-36   bug · ship    delivered    –         1h     ✓ PR #212
T-35   prototype     delivered    –         3h     ✓ branch
 
limits: max_running 2 · critical: src/core/** needs your sign-off

At a glance

For: Developers and small teams. Who already use coding agents.
Runs: In your terminal. On your laptop or an always-on machine.
Works with: Claude Code and git. GitHub and Codex next. Everything else through plugins.
Price: Free, open source core. At first release. Bring your own AI subscription.

What it does

Skelcrew turns work into tasks and runs your agents on them. A planner decides what kind of work it is, builders and testers do it, and you step in only when a decision is yours.

Fig 0.2

System

Work comes in

skel addGitHub issues soonYour tracker soon

Skelcrew

Triage→Build→Review→Gate

Planner · builder · tester · append-only log

You

Answer questions.

Sign off on what matters.

Agents in parallel

Claude Code today, Codex next, more through plugins

Work goes out

A branch, a pull request or a report

How a task moves

From one line of text to a pull request.

01 · Triage

T-41 · planner
intent  bug
rigor   ship
brief   VAT field loses its
        value on blur in the
        EU checkout form.
→ build

The planner decides whether it is a quick fix, a bug to investigate, a feature or an exploration, and how much care it needs.

02 · Build

T-41 · builder · wt/t-41
✓ failing test added
✓ fix in AddressForm
✓ 5 commits
✓ bun run check
 
→ review

A builder works in its own git worktree, test first, in parallel with every other task in flight.

03 · Review
```
T-41 · tester · round 1
changes requested
· empty VAT not handled
T-41 · tester · round 2
passed evidence: 3
 
→ gate
```
A separate tester reviews it adversarially and collects evidence. It passes the work or sends it back, up to three rounds.
04 · Done
```
T-41 · gate
touches src/core/**
? needs your sign-off
  y approve  n send back
 
✓ delivered PR #214
 
```
The task ends as a pull request. If it touches critical code, it waits for your sign-off first. You merge it, or send it back.

Why

You lead. Agents do the work.

Coding agents can do real work now, but each one still needs a person in a chat window, approving steps and reading every line. Run three at once and your day becomes switching between them.

That doesn't scale, and it wears people out. Skelcrew is built for the developer's well-being, not raw output. You act as the team lead who knows the architecture, and spend your attention on the decisions that matter.

It stays small and fits the tools you already use. No IDE, no new place to live. Its test for every feature: if the next model is twice as good, does this become dead code? If yes, it isn't built.

Why you can leave it running

Guarantees in code. Judgment in agents.

Agents propose. A small, deterministic core decides what's allowed, by fixed rules. An agent never holds the power to break them.

No agent approves its own work

Building and reviewing are separate roles. No agent can start more agents than you allow.

Nothing ships unchecked

The commit that passed review is exactly the commit that gets delivered. Critical paths wait for you.

Everything is on the record

Every step is an event in an append-only log. See what happened and why, and replay it.

Nothing is lost

Work is saved before an agent is stopped. A crash or restart picks up where it left off.

Fig 0.3

skelcrew.yaml

roles:
  planner: { harness: claude }
  builder: { harness: claude }
  tester:  { harness: claude }
 
repo:
  test: [bun run check]
  critical: [src/core/**]  # your sign-off
 
limits:
  max_running: 2

The whole setup for one repository.

Status

Skelcrew is being built by Skelcrew.

The rebuild works the way the product does: agents build it unattended, independent reviews by Claude and Codex check each step, and a human approves the decisions.

550+

automated tests on the core, including property tests that throw thousands of random scenarios at every rule

2×

fewer bugs from giving agents project rules and a spec, code, review loop, in a blind comparison

8 Oct

2026: the rebuild started, after an earlier version ran real agents end to end

doneThe core that decides what agents may do, and the store and loop that survive a crash
nowFirst real task end to end: skel add to a branch, with real agents
nextA terminal UI for what's running and what waits on you
nextCodex as a second harness, to prove Skelcrew isn't tied to one vendor
thenSkelcrew building itself, every day
laterHosted skelcrew.dev: steer your crew from anywhere, with approvals in the browser

Before you install

Questions, answered.

Is this an IDE?+

No. Skelcrew runs in your terminal next to the tools you already use. Each agent works in its own git worktree, and results arrive as branches and pull requests you review the usual way.

Does it merge code for me?+

Not on its own say-so. A separate tester reviews every change first, and anything touching paths you mark as critical waits for your sign-off. What gets delivered is exactly the commit that passed review.

Which agents does it work with?+

Claude Code today. Codex is next, through the same plugin contract, which keeps the core free of anything specific to one vendor.

Do I need a new account or AI subscription?+

No. You bring your own AI subscription, and the core, open source from the first release, runs free on your own machine. Skelcrew never resells model usage.

Will there be a hosted version?+

Yes, later. Hosted skelcrew.dev will let you steer your crew from anywhere, with notifications and approvals in the browser. The core stays open source.

When can I use it?+

When a real task reliably goes from skel add to a pull request. Leave your email and you'll hear the day it does.

Who's building it

Simon Carstensen

YCombinator S05

Senior full-stack developer with two decades of JavaScript and Python, based in Denmark. At Y Combinator in 2005 I co-founded Infogami with Aaron Swartz, and later co-created the original Jottit with him.

These days I focus on putting AI into the development process: agentic coding tools, MCP servers, and shaping stacks for agents to work in.

simonbc.com GitHub LinkedIn

Interested in backing Skelcrew? Get in touch and I'll send you the memo and a short demo.

Get updates

Be first to run a crew.

Leave your email and I'll write when there's something to try. One or two emails a year, no newsletter.