Your Claude Code agents run all day. Which half of what they ship is actually correct?

The ARC Engine Kit is the autonomy core that makes an agent stack loop until it converges, stay under a hard spend cap, and prove its work — so "it ran" finally means "it worked." Full Python source. Your API keys. No SaaS.

Get the Engine Kit — $299 See what's inside ↓

Built for Claude Code · runs anywhere Python does · 2 dependencies · 2026

"It runs" is not "it works."

Ships slop. No verification gate — so broken output reaches your users before it reaches you.

Leaks money. No cap. One loop bug = a four-figure API bill while you sleep.

Fires once and dies. Generate, exit. No retry, no convergence — the 80%-done result just sits there.

Lies. It reports "done." You can't verify, so you ship it. It rots until a customer finds it.

Not another pile of scripts. The tedious part you keep getting wrong.

You can wire up an agent in an afternoon. Making it converge, stay capped, and prove its work is the part that eats weekends and still breaks. That's these 16 engines. They drop into your Claude Code setup — not a framework you have to marry.

16 engines. 4 modules. 2 dependencies.

Loop `arcloop · queue_runner · boundary_policy`

Stops "fire once and die":

generate → gate → fix → re-run until it actually converges, then hands you an evidence report instead of a "done."

Safety `spend_cap · cost_gate · gemini_gate`

Stops the surprise bill:

hard daily + monthly caps, kill switch, shadow ledger. Ships at cap 0 — nothing auto-spends until you set a number.

Verification `proofcheck · completeness · multi_llm_check`

Stops slop shipping:

catches unreplaced {{templates}}, empty-shell output, and cross-checks an answer across Claude / GPT / Gemini before you trust it.

Reporting `notify · plainify`

Stops "what did it even do":

Discord alerts + plain-language rewrites of dense logs.

It proves itself on your machine — in 60 seconds.

$ pip install -r requirements.txt        # two packages
$ python tools/ops/arcloop.py --selftest
🎉 ALL SELFTESTS PASS
$ python tools/ops/spend_cap.py --selftest
✅ SELFTEST PASS  (daily cap = 0 until you set it)

No mockups, no "trust me." Unzip, run the selftests, read the source. These engines were pulled straight from a stack running several real businesses on Claude Code — which is also why there are no fake testimonials on this page. If you want proof, run the code.

One-time. Your keys. Your machine.

Start here

$299

Core. The 16-engine autonomy core. Full source, unlimited projects. Bring your own API keys.

Get Core

Everything

$499

Full. Everything in Core, plus expansion packs (research / content / sense), a setup walkthrough video, and a private Discord.

Get Full

No subscription. Updates: every new version of the engines you bought, free.

Questions

Do I need Claude Code?

Built for it, but most engines are plain Python and run anywhere with API keys.

Isn't this just scripts I could write myself?

You could. You won't — not the converging loop, the cap that actually holds, the gate that catches a lie. That's a weekend you don't get back, already built and selftested.

Refunds?

It's source code — once you download it, you have it, so no refunds. That's exactly why there's a free engine below and why this page lists what's inside down to the file names. Decide before you buy.

What about my API costs?

You run it on your own keys. spend_cap ships at 0 — nothing auto-spends until you set a number.

Can I resell it?

No. Use it in your own projects all you want; don't redistribute.

I'm not advanced — will it work?

If you already use Claude Code, you're past the hard part.

Try three engines free — no email wall.

spend_cap (the hard cap) · proofcheck (the slop-catcher) · boundary_policy (the auto-vs-ask rule). Run them, read them. If they're not your taste, you've lost nothing.

Get the Engine Kit — $299