v1.11.0 · Now with launch-readiness audit

The autopilot
that knows
when to stop.

For freelancers and teams in Claude Code whose bottleneck is review, not writing code. Give it a brief — it works issue by issue: plan, counter-review, implementation, tests, MR. You review over morning coffee and merge.

One-time · no subscription · 14-day money-back

Launch price until Aug 31 — €249 from Sept 1

superproject-auto — /projects/myapp

/superproject-auto --cycles 8

Backlog: 12 open · 3 WIP · 2 in review

── Cycle 1/8 ───────────────────────

✓ Issue #47: Auth middleware refactor

Plan → impl → tests → MR !108 opened

4 tests · lint clean · typecheck ok

── Cycle 2/8 ───────────────────────

✓ Issue #51: Fix payment webhook edge case

⊘ GATE: payment logic · awaiting sign-off

── Cycle 3/8 ───────────────────────

⧗ Conflict on MR !104 — rebasing

✓ Resolved · MR updated

3 MRs await your review. Stopping.

webhook note #42 → /superproject-respond · answered on thread (marker note=9001)

webhook issue #58 → /superproject-auto · debounced (identical trigger 4m ago)

// live run output

10 SKILLS · GITLAB · GITHUB · LOCAL MODE · ONE-TIME LICENSE · 14-DAY GUARANTEE · DUAL-LLM REVIEW · NO AUTO-RENEWAL · AUDIT TRAIL · WEBHOOK RELAY · SHIPCHECK · 10 SKILLS · GITLAB · GITHUB · LOCAL MODE · ONE-TIME LICENSE · 14-DAY GUARANTEE · DUAL-LLM REVIEW · NO AUTO-RENEWAL · AUDIT TRAIL · WEBHOOK RELAY · SHIPCHECK ·
00
Watch it work

See the autopilot
in action.

A full setup walkthrough — from install to your first automated MR.

01
The situation

More projects
than hands.
Each one stalls
the moment you
stop watching.

Maybe you've already tried AI agents on your projects. You know the pattern: excitement for the first hour, then a branch built on an unapproved MR, a "small improvement" nobody asked for, and in the worst case a half-wired payment integration — because "it was urgent."

The problem isn't that AI can't do the work. The problem is it doesn't know where to stop.

"Superproject is an autopilot built in reverse: boundaries first, throughput second."

02
How it works

From brief to done.
Ten skills.

/superproject

A brief or PRD becomes a roadmap, risks and issues with testable criteria. A vague brief doesn't pass — you get up to 5 clarifying questions, not a silent guess.

/superproject-sync

The backlog is pushed to GitLab/GitHub issues, idempotently and with labels. Adopt your existing backlog (--adopt) or pull the platform reality back (--reconcile).

/superproject-loop

Plan → counter-review → implementation → check against plan → tests + acceptance criteria → an MR/PR waiting for you. With Codex CLI, a second LLM reviews the same diff.

/superproject-auto

Bounded cycles, a concurrency lock, a ledger, notifications. Ideal in cron — you find finished MRs in the morning.

/superproject-add

A new request or bug: dedupe, triage, placement into the backlog — without replanning the whole project.

/superproject-status

What's happening, what's waiting on you, the single recommended next action. With --doctor it checks the whole environment.

/superproject-finish

Checks the finished project against the original brief, generates a changelog from the MRs/PRs, writes the final report. A project never just fizzles out.

/superproject-housekeeping

Runs on its own weekly clock, separate from feature delivery: layered security audit, dependency vulnerability scan, git and platform hygiene, stale-review nudges, allowlisted cleanup, relay health. Exactly one audit issue per run — and a fix branch only when a concrete safe fix exists. Never merges it.

/superproject-respond

A client writes a comment — the session answers on the thread where they asked. Triggered by the webhook relay with only numeric ids; it fetches the content itself from the platform API, answers questions, routes work requests through triage, and can start one bounded autopilot pass. Replies are idempotent: webhook redeliveries never double-post.

/superproject-shipcheck

Launch-readiness audit of a deployed app: nine plain-language checks, each proven broken before it's touched and proven fixed after.

New

Incoming GitLab/GitHub webhooks can feed new issues directly into a running session via the Webhook Relay — no polling, no glue scripts.

03
What it won't do

The guardrails
are the feature.

Not a disclaimer. Not a footnote. The five hard stops are the reason you can trust the autopilot with real client work.

01

It never merges.

Not even "merge once the pipeline is green." An unattended merge is still a merge. Every issue ends as an open MR/PR labelled workflow::review — and that's where the machine's authority stops.

02

Payments, auth and production need your sign-off.

Any issue touching money, login, migrations or production gets gated. Only your comment on that specific issue unblocks it. "The client said to just finish it" won't work.

03

Review agents are read-only.

They report findings, never edit. After every review the git status is checked — an unexpected change is reverted and logged. Every finding is a hypothesis, verified before it's fixed.

04

Verification that can actually fail.

No || true after a test, no PASS on a check the environment can't run. A small diff doesn't mean small risk — a behaviour change gets executed for real.

05

Bounded runs, an audit trail.

At most N cycles per run. When MRs waiting for review pile up, the run stops. Every run is written to a ledger — "why did this happen" always has an answer.

04
Day to day

What it looks like
in practice.

Finished MRs in the morning

A cron runs bounded cycles overnight. Every MR arrives with evidence: test output, passed acceptance criteria, a ledger entry. You just review and merge.

Merge conflict? Priority one

When an MR gets stuck on a conflict, the next run fixes it before touching anything else. A real collision is laid out for your decision instead of guessed.

New bug? One sentence

"Clients are complaining about confirmation emails" — /superproject-add dedupes, triages, respects the gates and queues it. No replanning the project.

A second pair of eyes from another LLM

With Codex CLI every diff passes a second model family. When models disagree, it verifies harder. A different model catches different failure modes.

The project learns from its mistakes

Every detected error pattern becomes a proposed lesson. Approve them once a week — the next runs get them baked into their prompts.

Alerts where you already are

A review-ready MR/PR reaches you via GitLab/GitHub natively. Optionally the autopilot sends a run summary to Telegram.

Client gets an answer in 60 seconds

A client writes "when will this be fixed?" in an issue comment at 15:40. At 15:41 the answer is on the thread — position in the backlog, what it's waiting on. If the comment was actually a new request, it's triaged into the backlog and the autopilot picks it up. You read the exchange in the evening.

05
New feature

Webhook Relay —
talk to your autopilot
from anywhere.

A webhook used to mean text waiting in the input for a human to press Enter. Now it can mean work. Map an event kind to a command, and an incoming issue or comment starts the autopilot or answers the client — while the event's text stays out of the prompt entirely.

01

GitLab & GitHub webhooks — now both for real

GitLab authenticates with a secret token header. GitHub can't send custom headers, so the relay verifies its HMAC signature (X-Hub-Signature-256) instead — the webhook secret is your relay token, compared in constant time.

02

REST API for everything else

curl, n8n, CI jobs: one authenticated endpoint to send text to any allowed session. Fire-and-forget; if Claude is mid-task, the prompt waits in the input.

03

Local cron

The same binary doubles as a CLI client. "Every Monday at 9:00, run maintenance" is one crontab line.

Built like the rest of Superproject: distrustful by default

  • Without autorun, webhook text is inserted without auto-submit — a human presses Enter. With autorun, only your own command is ever submitted.
  • One token per project, scoped to its sessions only. Constant-time comparison, 32+ char minimum, rotation is one config edit away.
  • Listens on localhost only; the sole public path is TLS through your reverse proxy. Never plaintext, not even during install.
  • Every rejected request is logged with its source address — journald is the audit trail.

Autorun: the webhook is a trigger, not a message

  • An event kind maps to an operator-written command: note → /superproject-respond, issue → /superproject-auto.
  • The event's prose (titles, descriptions, comments) never enters the submitted prompt. The only payload data allowed in are whitelisted placeholders — numeric ids and lowercase words, syntactically incapable of carrying instructions. Prompt injection has no channel.
  • The session fetches the real content itself from the platform API and treats it as data to evaluate, never instructions to follow.
  • Identical prompts are debounced per session; distinct events (two different client comments) both go through.
  • Headless-safe by default: with autorun configured, anything that doesn't resolve to a command is acknowledged and dropped — never left sitting in the input where the next Enter would submit it as part of an unrelated prompt.

One command to install — builds the binary, generates scoped tokens, sets up systemd and HTTPS:

bash

sudo ./install.sh -d hooks.example.com -p myproject:cl-myproject*

06
New feature
New in 1.11.0

/superproject-shipcheck —
is this actually ready
to point at real people?

You give it your live app URL and the repo. Nine plain-language checks run against the running app — each problem is proven real before anything is touched, a fix goes up on its own branch, and the fix is proven working on a throwaway preview build. One report, one verdict, and a fix branch per problem when it's safe to write one. It never merges — you always click merge.

01
BLOCKER

Who can read your data

Can a stranger with no login read private data, or is a master key exposed in the app?

02
BLOCKER

The two-account leak test

Can one logged-in user read another user's private data? The leak scanners can't find, because it needs two real accounts.

03
pass

The bill that lands on you

Can someone hammer an expensive endpoint with no brake and run up your bill? Also public forms with no bot protection.

04
not verified

The money path

Can someone get the paid product for free by forging a "payment succeeded" message, or set their own price?

05
pass

Can anyone find you

Share preview, robots.txt, sitemap, 404 page, HTTPS headers.

06
disabled

Can your AI be tricked

Prompt injection, leaked instructions, data the assistant shouldn't reveal. Says "no AI integration found" when the app has no AI.

07
BLOCKER

Can someone break into an account

User enumeration on signup and password reset, no brake on password guessing, one-time reset links that still work twice.

08
skipped

Does your app trust the browser

Validation that only runs client-side, error messages that hand over the database schema, and cross-origin access any website can use.

09
pass

Are you covered legally

Reachable privacy policy, where the data lives and who else touches it, consent before tracking, and copyleft licences pulled in by an assistant. Not legal advice.

It never merges. You decide what ships.

  • One verdict at the top of the report: 🔴 do not ship · 🟠 ship once these are merged · 🟢 safe to ship as far as these nine checks can see. At most five problems per run, ranked worst first, with a count of what was held back.
  • It proves each problem is real against the running app before it touches anything, then proves the fix worked on a throwaway preview build — instead of trusting comments or a README.
  • It never merges and never deploys — a fix branch is opened per problem when it's safe to write one, and the same read-only discipline applies until you click merge yourself.
  • Active tests run only against a URL you confirm you own, in-session, every run — nothing is stored, so nothing can probe unattended. Payments are only ever tested in Stripe test mode, and the account tests only touch a throwaway address the skill creates and removes itself.
  • Every report closes with a map onto AIUC-1, the emerging certification standard for AI agents. A snapshot, not a certification — real certification needs an accredited auditor and quarterly adversarial retesting.
07
Pricing

One payment.
Everything included.

No subscription. No feature tiers. No surprise at month 2.

Launch pricing ends Aug 31. From Sept 1: Individual €249 · Studio €749.

Individual

€129one-time

Launch price until Aug 31 · €249 from Sept 1

  • All 10 skills: /superproject through /superproject-shipcheck
  • Webhook Relay included
  • Full documentation (12 chapters), security model, operational guides
  • Config templates and a lessons loop that gets smarter each week
  • Local mode: the full cycle on plain git, no platform account needed
  • 12 months of updates from the private marketplace
  • Email support from me — not a chatbot
Buy Individual License →

Studio

€429one-time · 5 seats

Launch price until Aug 31 · €749 from Sept 1

Everything in Individual, for the whole team. One purchase, shared config, consistent workflow across all projects.

Need more seats? Email me →

Buy Studio License →
14-day money-back guarantee · no questions asked
08
Questions

Honest answers.

// ready to ship

Stop watching.
Start reviewing.

One license. Every future project. You show up to morning coffee and find MRs waiting — not chaos.

Secure checkout via Stripe · 14-day money-back

Launch pricing ends Aug 31 — €249 / €749 from Sept 1

Made with AI in Macaly