Most founders choose an AI builder for speed and taste: one feels faster, one writes cleaner code, one deploys with a click. Few choose it for what it means legally, and that is reasonable, because the law does not care which tool you used.
What does change from tool to tool is where the legal mistakes creep in. A browser builder that hosts your app makes different defaults from a code agent working inside your own repository. Knowing the pattern for your tool tells you where to look first.
This guide groups the six tools people ask me about most into two families, explains the gaps each family tends to leave, and ends with a short checklist that works for all of them. For a deeper, tool-by-tool walkthrough with ready-made fix prompts, I link to the SafeLaunch guide for each one.
Key takeaways
- Privacy, cookie, consumer and accessibility laws apply to what the app does; no regulator asks which tool wrote it.
- Browser builders (Bolt, v0, Replit) make the first deploy public, so a test build can collect real people’s data on day one.
- Code agents (Cursor, Claude Code, Windsurf) work inside your own stack and machine, so secrets, production credentials and dependencies need the most care.
- Every tool reads a project instructions file or first message; a short set of rules there prevents most common gaps.
- Test the live app in a private window as a logged-out user and as a second account; that one check finds the most serious problems.
Does the tool I build with change which laws apply?
No. Data protection law asks who decides why and how personal data is used. If you launched the app, that is you, so you carry the main duties whether you typed every line or none. Consumer law asks who sells to the public, and again that is you. Accessibility law asks whether disabled people can use the service you offer.
The AI tool is a supplier. Its terms almost always say that you are responsible for what you build, configure and deploy. That holds for Bolt, v0, Replit, Cursor, Claude Code and Windsurf alike.
So the useful question is not ‘is this tool compliant?’ but ‘what did this tool decide for me, and what did it leave to me?’ The answer depends on which family the tool belongs to.
| Family | Tools | What it decides for you | Where gaps usually appear |
|---|---|---|---|
| Browser builders | Bolt, v0, Replit (and Lovable) | Framework, hosting, often the database and sign-in | Public deploys, database access rules, trackers in the page head |
| Code agents | Cursor, Claude Code, Windsurf | Very little; it works in the stack you already have | Secrets, production access, dependencies, rules never written down |
Browser builders: Bolt, v0 and Replit
Bolt (from StackBlitz), v0 (from Vercel) and Replit build and run the app in your browser and put it online with one click. Bolt projects often use Supabase for data and sign-in. v0 builds Next.js apps that deploy to Vercel, with integrations for databases such as Supabase and Neon. Replit’s Agent can use Replit’s own database, sign-in and hosting.
That convenience creates one recurring problem: the line between a test and a launch disappears. A shared preview link is a public website. If it has a sign-up form, it is collecting personal data, and every duty that comes with that applies from the first email address.
The second problem is the database. The builder creates tables quickly, but access rules, such as Supabase’s row-level security, decide who can read and write each row. If they are missing or too broad, anyone with the public key in your page can read other users’ records. A security researcher showed exactly this across many Lovable apps in 2025, and the same mistake is easy to make in any builder that wires up a database for you.
The third is tracking. Ask a builder to ‘add Google Analytics’ and it will usually paste the tag into the page head, where it fires on the first page load. In the EU, and for most tags in the UK, that tracker needs consent before it runs, so a banner added later does not fix it.
- Treat the first public deploy as your launch date for privacy purposes.
- Turn on database access rules for every table and test them before inviting anyone.
- Load analytics and ad tags only after consent, through your consent banner rather than the page head.
- Keep payment, database service and AI API secret keys in server functions, never in the browser bundle.
Tool-by-tool fixes: Bolt, v0 and Replit compliance guides on SafeLaunch.
Code agents: Cursor, Claude Code and Windsurf
Cursor and Windsurf are code editors with an AI agent built in. Claude Code is Anthropic’s agent that works in your terminal and repository. None of them hosts your app. They work inside whatever stack you already have, read and edit files, and run commands on your machine.
That shifts the risk. The agent will not quietly pick a database or publish a preview, but it can read anything in your project folder, including environment files. If your local setup points at the live database, the agent is one wrong command away from changing real users’ data. In July 2025 the SaaStr founder reported that Replit’s agent deleted a live database during a test, and the lesson applies to every agent with production access.
The second risk is dependencies. Agents add packages freely. Most are fine, but each one is code you now ship, with a licence you must follow and vulnerabilities you must patch. A copyleft licence such as the AGPL can require you to publish your own source code in some cases.
The third is that the rules live in your head. A code agent follows instructions well, but it only knows the ones you write down. Without them, it will add a tracker, a form field or a third-party API without asking whether your privacy notice covers it.
- Give the agent a development database with test data; keep production credentials off your laptop where you can.
- Keep secrets in environment variables that are excluded from version control, and rotate any key that was ever committed.
- Run a dependency and licence check before each release, not once at launch.
- Review the agent’s changes as you would a new developer’s: read the diff before you merge it.
Tool-by-tool fixes: Cursor, Claude Code and Windsurf compliance guides on SafeLaunch.
Where do I put compliance rules so the AI follows them?
Every one of these tools reads instructions before it works. Cursor reads project rules from the .cursor/rules folder. Claude Code reads a CLAUDE.md file at the start of every session. Windsurf has workspace rules. Bolt and v0 offer project instructions on some plans, and Replit’s Agent will follow a project file you point it to. AGENTS.md is an open format that a growing number of agents read too.
A rules file will not make an app compliant on its own, and it does not replace checking the result. But it stops the most common gaps being created in the first place, which is far cheaper than finding them later.
Keep the rules short and concrete. An agent follows ‘never load analytics before the visitor accepts cookies’ far better than ‘be GDPR compliant’.
| Tool | Where the rules go |
|---|---|
| Cursor | A rule file in .cursor/rules (older projects: .cursorrules) |
| Claude Code | CLAUDE.md in the repository root |
| Windsurf | Workspace rules in the project |
| Bolt | Project instructions, where your plan offers them, or the first message |
| v0 | Project instructions, or the first message of the chat |
| Replit | A project file such as replit.md or AGENTS.md that the Agent is asked to follow |
The gaps every AI-built app shares
Whichever tool you use, the same handful of problems turn up again and again. They are not exotic legal questions; they are defaults nobody revisited.
A privacy notice generated from a template describes a typical app, not yours. It names vendors you do not use and misses the ones you do: your database host, email provider, payment processor and AI model provider. A notice that says something untrue is a problem in its own right.
AI features need a disclosure. Since 2 August 2026, Article 50 of the EU AI Act requires that people in the EU are told when they are talking to an AI system, unless that is obvious. Several US states have similar rules for particular uses.
Forms and buttons are often unusable with a keyboard or screen reader: placeholder text instead of labels, custom menus that trap focus, low-contrast buttons. That is the most common basis for website accessibility claims in the US and matters under the European Accessibility Act.
How do I check an AI-built app before launch?
Start with the test that finds the most serious problems in five minutes. Open the live app in a private browser window. Without logging in, try to load data that should be private. Then create two test accounts and try to see the first account’s data from the second. If either works, stop and fix the access rules before anything else.
Next, reload the home page in a fresh private window and open the browser’s developer tools. Before you click anything on the cookie banner, look at the cookies and network requests. Analytics or advertising requests at that point mean trackers are firing before consent. Click Reject in a new session and check again.
Then compare your privacy notice with reality: list every service the app sends data to and make sure each one is named. Finally, tab through your sign-up and checkout forms using only the keyboard.
Want the fixes written as prompts for your tool? See compliance fix prompts for Lovable and Cursor, or take the free 2-minute app self-check.
Checklist
- Decide which family your tool belongs to and read the matching section above.
- Write five or more concrete compliance rules into the instructions file your tool reads.
- Turn on database access rules for every table and test them as a logged-out user and a second account.
- Move every secret key to server-side code and rotate any key that has been exposed.
- Keep production credentials away from your coding agent; give it a development database.
- Load analytics and ad tags only after consent, and test Reject in a fresh session.
- Rewrite the privacy notice so it names every service that receives user data.
- Add an AI disclosure to any chatbot or AI-generated content shown to users in the EU.
- Tab through sign-up and checkout with the keyboard and fix anything you cannot reach.
FAQs
Is one AI app builder more legally compliant than another?
Not in a way that protects you. Some have better defaults or built-in security scans, which help, but the legal duties attach to the app and the business running it. A careful founder on any tool beats a careless one on the ‘best’ tool.
Is a Bolt or v0 preview link a launch?
For privacy purposes, yes, once real people can use it. A public link with a working sign-up form collects personal data, so the privacy notice, consent and database rules must already be in place.
Can I ask Cursor or Claude Code to make my app GDPR compliant?
You can ask it to fix specific, testable things: block trackers until consent, add labels to fields, move keys to the server. ‘Make it GDPR compliant’ is too vague for an agent and too broad for code alone, because some duties, such as contracts with processors, sit outside the codebase.
Does a rules file like CLAUDE.md or .cursor/rules make my app compliant?
No. It makes the agent less likely to create common problems. You still need to check the deployed app, keep your privacy notice accurate and handle the duties that are not code, such as breach plans and data processing agreements.
Do I need to tell users my app was built with AI?
Generally no. Disclosure rules apply to AI that users interact with or AI-generated content they see, such as a chatbot or realistic images, not to the tools used to write the code.
Who is responsible if the AI agent deletes or leaks user data?
You are, as the business running the app. You may have a claim against a provider if its own platform failed, but you still owe any breach notifications to regulators and users, often within 72 hours in the EU and UK.
Primary sources
General Data Protection Regulation (EU) 2016/679, Articles 25 and 32ePrivacy Directive 2002/58/EC, Article 5(3) (consent for cookies and trackers)EU AI Act, Regulation (EU) 2024/1689, Article 50 (transparency)Cursor documentation: project rulesClaude Code documentation: CLAUDE.md project memoryAGENTS.md: the shared instructions format for coding agentsGeneral information, not legal advice.
Go deeper on SafeLaunch
Cursor app compliance: GDPR, cookies and securityClaude Code app compliance: GDPR, cookies and securityWindsurf app compliance: GDPR, cookies and securityBolt.new app compliance: GDPR, cookies and securityv0 app compliance: GDPR, cookies and securityReplit app compliance: GDPR, cookies and securityCompliance fix prompts for Lovable and Cursor