Seaworthy is a pre-launch security check for apps built with AI coding tools such as Cursor, Claude Code, Lovable and Bolt. It is aimed at the people shipping that code — solo builders, indie developers and agencies — and it exists to answer one question before launch: is this app safe to put in front of real users? You run it with a single command, npx seaworthy-cli, inside your project folder. There is no account to create and nothing to install beyond Node 20 or later. Seaworthy reads your project where it sits, works out which stack you are using, and runs only the checks that apply to it.
The problem it addresses is a specific side effect of fast AI-assisted development. Cursor, Claude Code, Lovable and Bolt write working code fast — but the site states plainly that they also skip authorisation, leak keys into the browser and leave database tables open to anyone with your public key. Those are the gaps that do not surface when the app appears to work. A page renders, a form submits, a payment completes, and underneath a table is readable by anyone holding the public anon key, or a webhook accepts a request that never came from Stripe. Seaworthy is built to find those gaps on your own machine, before your users do, rather than after an incident forces you to go looking for them.
The secrets checks cover the many ways a key ends up somewhere it should not be. Seaworthy looks for API keys and passwords committed to the repository, both in the current code and in its git history, because deleting a key from a file does not remove it from the commits behind it. It flags .env files that are committed or not git-ignored, secrets given a public prefix so they ship to the browser, and LLM keys used in browser code. Every report masks secrets, so a finding can be shared with a client, an investor or a teammate without handing over the very credential the check is warning about. Each finding comes with the file and line, what an attacker could do, how to fix it, and a prompt you can paste into your AI coding tool to have the fix written for you.
The database and storage checks target the open-by-default settings that AI tools tend to leave behind. For Supabase, Seaworthy flags tables with no row-level security, policies that let everyone in (the documentation shows the pattern as using (true)), and views or security definer functions that bypass row-level security altogether. It also flags public Supabase buckets holding private files. For Firebase it looks at open Firestore, Realtime Database and Storage rules, including test mode that was never switched off. These are the findings most likely to turn into a data exposure, because the database is reachable by design and often needs only a public key that already ships in the client bundle.
The remaining checks cover API routes, payments, AI agents and dependencies. On the API and payments side Seaworthy flags routes and server actions with no auth check, records fetched by ID without checking the owner, missing rate limits on AI, email, SMS or login endpoints, CORS that lets any site make credentialed requests, and SQL built from user input. It also catches Stripe webhooks that do not check the signature and prices taken from the client. On the AI and agent side it flags user input pasted into the system prompt, model output rendered as HTML or run as code, agent tools with shell or file access and no guard, MCP configs with plain-text secrets or open servers, and MCP servers auto-approved in Claude Code settings or launched through a shell. Finally it reports dependencies with known vulnerabilities from OSV, packages that do not exist on npm (often invented by AI tools) or that are brand new, and missing security headers.
How the tool works is deliberately narrow: one command, nothing uploaded. Seaworthy reads your project where it sits, works out your stack, and runs only the checks that apply. Checks are written for Next.js, React, Vue, Svelte, Supabase, Firebase, Stripe and the major LLM SDKs, and each one is tuned for few false alarms by looking at what your code actually does rather than just at words in it. Findings carry a severity and a rule ID, the file and line with a snippet of the offending code, a Why explaining what an attacker could do, a Fix showing the corrected version, and a paste-ready prompt for your AI assistant. You can ignore a finding with a comment, but only by writing a reason. Output is available for the terminal and as JSON, with exit codes for CI, plus SARIF and Markdown on the free plan. There is also a local MCP server and a check-before-deploying step you can add to your AI assistant's rules file.
Because Seaworthy runs locally, your code, findings and file names never leave your machine. The only outbound traffic described is package names and versions sent to osv.dev for the dependency checks and package names sent to the npm registry, and the --offline flag skips both. With a paid plan the tool also sends your licence key, a machine name and IDs from the licence server, at most once a day — never code, findings or file names. A confirmed licence keeps working for 14 days offline, and in CI you set the SEAWORTHY_LICENCE secret, which does not use up a machine. The site is equally clear about limits: Seaworthy is an automated check that catches common, serious mistakes, not a certification or a guarantee, and no tool finds everything.
The practical benefit is timing. Catching an open table or an unverified webhook before launch costs a code change; catching it after launch can cost a data exposure. Because every check is free and the tool needs no account, it can be run on every project without a procurement conversation, and because the output includes both fix text and a paste-ready prompt, the loop from finding to corrected code stays inside the same AI workflow that produced the code. Anyone who has shipped an AI-generated app will recognise the pattern the tool describes: the code works, but nobody wrote the authorisation check.
Typical use is a final pass over a project before it goes live, run from the project folder with a single command. In CI, the free plan provides a GitHub Actions workflow in one command, and exit codes let a build fail when a serious finding appears. On Pro, pull-request checks post one comment listing only the problems a change introduces, weekly scans alert through Slack, a webhook or email when something new appears, and scan history with before/after comparison shows whether the codebase is getting better or worse. Agencies can scan every project in a folder into one combined report and deliver self-contained HTML reports, white-labelled with their own logo, agency name and client name.
Pricing separates the checks from the workflow around them. Free costs £0 and includes every check, terminal, JSON, SARIF and Markdown output, exit codes for CI, a one-command GitHub Actions workflow, fix text and paste-ready AI prompts. Pro is £99 per year or £12 per month for solo builders and adds pull-request checks, weekly scans with Slack, webhook or email alerts, a self-contained HTML report to share with clients or investors, scan history and trends, before/after comparison and since-last-scan results, priority support and 2 machines per licence key. Agency is £349 per year and adds white-label reports, folder-wide scanning into one combined report and 10 machines per licence key. A paid feature never blocks a scan: without a plan, Seaworthy still runs and tells you which plan adds it. Checkout is handled by Polar, and after buying you take the key from the customer portal and run npx seaworthy-cli login with it. The source is available on GitHub under the Functional Source License, with each release becoming Apache 2.0 after two years.
Seaworthy's value proposition is a single local command that closes the security gap AI coding tools leave behind, gives you the file, line, explanation and fix for each problem, and keeps your code on your own machine the entire time.