Stop paying for Lovable, Bolt, v0 and Replit. Hosted builders sell the setup around the AI — rules, design and backend conventions, connection and go-live gates. Agentic Ship is that setup as a toolkit you own and drop into your own workspace, driven by the coding agent you already pay for. No app is bundled to delete — the kit directs and verifies the one your agent builds.
Run it from the project directory you want to adopt it in:
npx github:moasq/agentic-shipKeep the github: prefix — it ships from GitHub, not npm. That copies the toolkit into
the current folder (existing files are skipped, never clobbered); then pnpm install and
open the folder in your agent. That's it. Say what you want to build; it reads
AGENTS.md and follows the rules there. Flags: --force overwrites,
--merge folds the pnpm scripts into an existing package.json, --dry-run previews.
As a plugin. In Claude Code or Codex the same repo installs as a plugin — canonical skills and the pinned MCP catalog, one copy of every rule:
/plugin marketplace add moasq/agentic-ship
/plugin install agentic-ship@agentic-ship
From source, to develop the toolkit itself:
git clone https://github.com/moasq/agentic-ship.git
cd agentic-ship && pnpm install && pnpm verifyEvery layer is chosen, pinned in skills.lock.json, and verified by gates. The kit's job is making an agent use them correctly together. The full reasoning behind each pick lives in docs/stack.md.
| Layer | Choice |
|---|---|
| Runtime |   |
| Framework |    |
| Styling |  globals.css, no config file |
| Components |   |
| State | RSC props first, then URL state, then Zustand 5 |
| Backend |  |
| Auth |  |
| Billing |  @convex-dev/stripe — hosted checkout, webhook-backed entitlement |
 @convex-dev/resend — test-mode by default |
|
| Analytics |  /ingest proxy; the CSP stays closed |
| Fonts | Self-hosted OFL faces, fetched by pnpm font, committed |
| Gates |  pnpm verify on every completion, pnpm verify:full before a release |
| UI gates | pnpm ui:plan direction contract · pnpm ui:review visual evidence · pnpm check:ui component boundaries |
| Deploy |  netlify.toml is authoritative |
| Delivery |  gh CLI device-flow OAuth for repo, PRs, and CI |
| Tracking |  |
| AI hosts | |
| Tool catalog | 11 pinned MCP servers in .mcp.json, mirrored per host |
| Writing | Vendored docs handbook, humanizer pass, README doctrine, plain-language audit, ADRs — docs are gated prose |
Stack marks are vendored, not hotlinked — provenance in .github/assets/stack/credits.md. Zustand and Playwright ship no mark in that set, so their rows stay text-only.
- ✅ Pinned — every version, registry, and MCP server locked in skills.lock.json
- ✅ Gated —
pnpm verifyis the definition of done on every completion, not a release ritual - ✅ Consent-first — services connect through the vendor's own OAuth, and every connection is revocable
- ✅ Offline-first — a fresh copy verifies green with nothing connected
Opinionated does not mean stuck. Every provider sits behind a seam you own, so the honest question is not "is it supported" but how much code a swap touches:
| Want instead | Wired today | Swap cost |
|---|---|---|
| Vercel, Cloudflare | Netlify | small — the deploy seam is netlify.toml, one build command, one doc; product code never names the host |
| Plausible, Umami | PostHog | small — src/lib/analytics.ts is the only file that imports the SDK |
| Postmark, SendGrid | Resend | medium — convex/email.ts is the only sender, but the Convex component and its webhook go with it |
| Polar, Lemon Squeezy | Stripe | medium — checkout and portal are one seam; entitlement rides the @convex-dev/stripe component |
| Clerk, Auth.js | Better Auth | medium — session truth is one query behind requireUser, but the Convex adapter is load-bearing |
| Supabase, Postgres + Prisma | Convex | large — Convex is the spine; the auth, billing, and email components all ride it. Swapping it means rebuilding those seams |
That last row is this kit's own lock-in, stated plainly. The difference from a hosted builder: all of it is MIT-licensed code in your repo, so the exit is a refactor you can run, not an export you buy.
Hosted builders are real products, and the running app they put in a browser tab within minutes is a real advantage. The table shows what that convenience costs, and what a toolkit in your own repo does that a rented builder cannot.
Compared against the shipping versions as of August 2026: Lovable with Lovable Cloud and Build mode, Bolt V2 with Bolt Cloud, the rebuilt v0 platform (February 2026), Base44 under Wix, and Replit Agent 3. Where they improved, the table says so. Marks come from each vendor's public docs and product pages; if a cell is wrong or stale, open an issue — the table only works if it stays true.
✅ yes ·
| Agentic Ship | Lovable | Bolt.new | v0 | Base44 | Replit | |
|---|---|---|---|---|---|---|
| Open source | ✅ MIT | ❌ | ❌ | ❌ | ❌ | ❌ |
| Cost | ✅ $0 forever — MIT, drives the agent subscription you already have | |||||
| Bring your own coding agent | ✅ any of nine hosts, swap anytime | ❌ built-in agent | ❌ built-in agent and model | ❌ built-in agent | ❌ built-in agent | |
| Code lives in your git repo | ✅ from the first commit, no platform copy | ✅ two-way GitHub sync | ✅ native GitHub: branches, PRs, existing repos | ❌ one-way push, on a paid tier | ||
| Take everything with you when you leave | ✅ nothing to leave; the repo is the product | ❌ front end exports on a paid tier; the backend stays behind their SDK | ||||
| Self-host anywhere | ✅ any Node host | ❌ their runtime only | ||||
| Backend built from parts you can hire for | ✅ Convex · Better Auth · Stripe, named and pinned in your repo | ❌ proprietary | ||||
| Definition of done you own | ✅ pnpm verify in your repo and CI: contracts, tests, visual evidence |
❌ checks run in their pipeline | ❌ checks run in their pipeline | ❌ checks run in their pipeline | ❌ checks run in their pipeline | |
| Security gates before going live | ✅ dependency audit, closed CSP, production preflight | ✅ auto scan on publish, deep scan on demand | ❌ | ❌ | ❌ | ❌ |
| Enforced design direction per product | ✅ written plan, tokens, review gates | ❌ | ❌ | ❌ | ||
| Secrets stay between you and the vendor | ✅ vendor OAuth, hidden terminal input | |||||
| Zero install, runs in a browser | ❌ needs a terminal | ✅ | ✅ | ✅ | ✅ | ✅ |
| Shareable hosted preview in minutes | ❌ you deploy it | ✅ | ✅ | ✅ | ✅ | ✅ |
| Built for non-developers | ❌ assumes a coding agent | ✅ | ✅ | ✅ | ✅ | ✅ |
| Native mobile output | ❌ web only | ❌ web only | ✅ Expo | ❌ web only | ✅ iOS builds |
Two fairness notes on cost. The free tiers are real, and fine for a weekend test — but building a product burns through them, and that is where the paid plans begin. And hosted credits include the model usage, so a builder is only a second subscription when an agent subscription already exists — which is exactly who this kit is for. The last four rows are a real recommendation: if you will never open a terminal, or you want a phone app from a single prompt, a hosted builder is the better buy. This kit is for the other case — a coding agent is already open, and the product has to outlive the demo with owned code, a portable backend, and a definition of done you can run yourself.
Meridian, an uptime monitor, shipped end to end from this kit: Convex backend, Better Auth sessions, Stripe plans, Netlify deploy. The kit bundles no UI of its own — every pixel below was produced by a coding agent working under the kit's rules and gates.
Claude Code, Codex (plugins + native adapters), Cursor (generated agents, MCP mirror, hooks), Hermes (installable profile), and OpenClaw (repo as agent workspace). Windsurf, Cline, Copilot and Gemini CLI read AGENTS.md and the plain-markdown skills directly. One authored rule set, generated adapters per host.
Host marks above are vendored, not hotlinked — provenance in .github/assets/hosts/credits.md.
The wiki lives in docs/ — one article per question:
- What is the Agentic Ship stack? — every layer and why it was picked
- How do I go from empty folder to shipped product?
- How is the toolkit put together? — one rule, one home
- How do service connections work? — consent, receipts, revocation
- How do I see what the agents are doing? — the work queue, Linear, GitHub
- How are these docs written? — the writing skills behind the prose
Reference material:
- AGENTS.md — every rule and every
pnpmcommand, in one file - .agents/skills/ — procedures: product lifecycle, connections, UI, backend, security, testing, launch
- .agents/agents/ — five specialist role briefs, plus the vendor-generated Playwright trio
- Visual direction — the planning procedure in .agents/skills/visual-direction/SKILL.md, its anti-slop rubric, reference sources and real-site gallery, with .agents/ui/plan.example.json as the worked plan
- Visual review — the acceptance policy in .agents/skills/visual-qa/references/review-policy.md