AI App Builder Custom Code in 2026: Can You Edit It, and Does Your Edit Survive?

Quick Answer (September 2026). Seven of the eight builders we track let you edit your app's code in some form, but "can I edit it" is the easy half of the question. The half that bites is what happens next. Only Bolt.new publishes a way to tell the AI which files to leave alone. Only v0 by Vercel documents that your manual edits do not create a restore point, while Lovable documents the exact opposite, that every manual save becomes a revertable version. Lovable gates code editing behind a paid plan and Softr gates its custom code block the same way, while Bolt.new turns its editor read-only if you happen to be using Safari. Bubble has no application source code to edit at all, so its hatch is an insert point rather than an editor. Check the recovery story before the editor, because the editor is the part every vendor advertises and recovery is the part that decides whether an afternoon of manual work survives your next prompt.

"Can I edit the code" is four questions wearing one name

Ask a vendor whether you can edit the generated code and the answer is almost always yes. That answer is close to useless, because it collapses four independent things that fail separately.

1. Is there an editor at all? Can you see and change the files from inside the product.
2. Are you allowed to write in it? An editor you can read but not save to is a viewer. This is where plan gates and, on one builder, browser gates live.
3. Does your edit survive the next prompt? You hand-edit a component, then ask the agent for one more feature. Does your work still exist afterwards.
4. If it does not survive, can you get it back? A manual edit is only as safe as the restore point it created, and one builder here documents that it creates none.

Almost every feature table published on this subject answers question one and stops. The differences that cost people an afternoon live in questions three and four, and the vendors document them very unevenly. Below is what each vendor's own documentation says, read in September 2026.

The code escape hatch across eight AI app builders, September 2026

Every cell is drawn from the vendor's own current documentation. Where a cell reads "not documented", that is an absence in the documentation rather than a proven absence in the product.

Finding one: the gate on writing is not always a price

Lovable is explicit that the editor is not fully yours on the entry tier. Its documentation states that code editing is available on paid plans and that on the Free plan the editor is read-only, showing a "Read only" label and an upgrade option. Its own FAQ repeats it: opening the editor needs edit access to the project, and editing and saving additionally requires a paid plan.

What makes this more interesting than an ordinary paywall is the asymmetry sitting next to it. Lovable's Git sync is available on all plans. So on the Free tier you cannot change a line in the browser, but you can connect a repository and change whatever you like in your own editor. The free-tier escape hatch is not a smaller hatch, it is a different door, and it leads out of the building.

Softr gates in the conventional way. Its documentation states that custom code is available for users on the Basic plan and above, and repeats that you need to be on a paid plan to use the Custom code block.

Bolt.new has the most unusual gate in the category, and it has nothing to do with money. Its documentation carries a note that if you are using Bolt in Safari, Code view is read-only: you can open and look through your files, but you have to prompt Bolt to make changes instead of editing them directly. A capability that depends on your browser choice is not something a pricing page will ever tell you, and it is the kind of detail that only turns up when you are already committed.

Finding two: one builder of eight lets you tell the AI to keep its hands off

This is the single sharpest split we found, and it is the one that answers question three.

Bolt.new publishes two complementary controls in Code view. You can right-click files and choose Target file to limit Bolt to specific files, and you can right-click a file or directory and choose Lock file for a single file or Lock all for a directory to exclude them. That is a documented, user-operated boundary between your hand-written work and the agent's next pass. No other builder in this comparison documents an equivalent.

The absence elsewhere is not an accusation, and it is worth being precise about what it means. It means that on the other seven, the published answer to "will the agent rewrite the file I just edited" is that there is no published answer. In practice people manage this by prompting carefully and by keeping changes small, which works until the change is neither small nor easy to describe.

Base44 approaches the same anxiety from the opposite direction. Rather than letting you fence the AI out, it lets you inspect what the AI did before any of it goes live. Its Code changes modal shows every file the AI touched, grouped by folder, with a side-by-side diff, a stats bar counting files added, modified and deleted, and colour coding: green for newly added, amber for modified, red for deleted. Its documentation also warns that saving code is not the same as shipping it, and that you must click Publish for changes to go live. That separation means a surprising AI edit has two chances to be caught rather than none.

Finding three: the two answers to "can I get it back" are opposites

v0 by Vercel has the most capable editor in the group on paper. Its documentation describes syntax highlighting, global search across files, a diff view for inspecting each generation, split view for editing two files side by side, and full file creation, renaming and deletion. If you were scoring question one alone, v0 would win.

Then read its versions page. It states that each time v0 updates a code block from a message it creates a new version, and that non-message actions, such as editing code or modifying files directly, do not generate new versions. Read that carefully, because it is the difference between a safe afternoon and a lost one. On v0 the agent's work is versioned and your work is not. There is no checkpoint corresponding to the state you hand-built, so "restore the version from before it went wrong" restores a version that never contained your edit.

Lovable documents the exact inverse. Each save from the code editor creates a version in version history, and its FAQ says manual edits can be reviewed or reverted the same way as Lovable's own changes. It is equally blunt about the risk that makes this matter: Lovable does not review your manual edits, so if a change breaks something you revert it from version history or ask Lovable to fix it. The safety net is explicitly the version, not any validation.

Replit is the most conventional answer and, for anyone who already uses Git, the strongest. Its documentation describes four version control surfaces sitting on one underlying Git repository: Agent Checkpoints that commit automatically at milestones, a Git Pane for visual access, a Git CLI offering command-line access to all Git functionality, and File History for granular changes inside individual files. Its own comparison table is candid about the trade: rollback is one-click for Agent Checkpoints and File History, and "Requires Git knowledge" for Git commits.

Two products, the same question, opposite answers, and only one of the eight gives you a switch. If you intend to hand-edit at all, this is the cell to read first.

Finding four: where there is no code, the hatch is an insert point

Bubble belongs in this comparison for the opposite reason to everyone else. There is no application source code to open, so questions two, three and four do not arise. Its documentation describes the escape hatch as writing custom JavaScript using the HTML element or by creating custom plugins, which it says enables importing and using JavaScript modules or packages directly within a Bubble application. You are not editing what Bubble generated. You are adding your own code alongside it, through a defined slot.

Softr works the same way and publishes a caveat worth more than most feature bullets. Its Custom code block inserts HTML, CSS or JavaScript into the body, with application-level and page-level settings for the head. Because Softr apps run as single-page applications, its documentation states that scripts in the app header or footer execute only once, on the first page load, and will not re-execute when navigating between pages, unlike in non-SPA mode. It also warns that elements injected by your JavaScript are not cleaned up on navigation, and points at a window.SOFTR_PAGE.beforeUnload API to remove them. A snippet that works perfectly on a hard refresh and silently stops working when a user clicks through your nav is a genuinely hard bug to find.

Webflow sits at the far end. Its code components let you build or import React components for use inside and outside Webflow, which is real code ownership, but the constraints are published and they are firm: React Server Components are not supported, you must explicitly import external styles such as site variables and tag selectors, and because all JavaScript runs in the browser and is visible to users you should use public APIs only and never include secrets or sensitive keys. It also requires the API you call to accept cross-origin requests. This is a code path that assumes you have a build toolchain and a repository, which is a different proposition from clicking a Code tab.

What to check before you commit

Owning the code and being allowed to edit it are separate questions, and they are answered in different places. If what you care about is whether the source is yours to take away, that is covered in our comparison of code ownership across AI app builders. This page is about the narrower and more immediate question of what you can do inside the product today.

Four checks, in the order they will actually hurt you:

• Check the recovery story before the editor. Find the vendor's versions or history page and look for whether a manual save creates a restore point. If the documentation does not say, assume it does not.
• Check the write gate on the tier you will actually buy. A read-only editor on the free plan is easy to miss during evaluation, because during evaluation you are mostly prompting.
• Check whether there is any way to fence the AI out. If there is not, plan to keep your manual edits in a repository rather than in the product.
• Check whether an insert point is enough. If the builder has no source code, the question is not how good its editor is but whether a slot for your own HTML and JavaScript covers what you need.

The two builders whose models sit furthest apart on these four checks are Lovable and Bolt.new, and we compare them directly in Lovable versus Bolt.

Verdict, September 2026

If you expect to hand-edit regularly and want the AI to respect your work, Bolt.new is the only builder here that publishes a keep-out mechanism, and Lock file is the feature we would miss most on the other seven. Check Safari first.

If you want manual edits treated as first-class and recoverable, Lovable documents the clearest answer, with the caveat that you have to be on a paid plan to write anything in the browser at all.

If you already work in Git, Replit is the least surprising, because the escape hatch is not a product feature but the repository underneath it.

If you want to see what the AI changed before it ships, Base44 has the best-documented review step in the category.

v0 by Vercel has the best editor and the weakest recovery guarantee, which is an unusual combination and the reason it needs naming rather than scoring. Use it, and keep your own copy.

Bubble, Softr and Webflow are not worse on this axis so much as differently shaped. They are answering "can I add my own code" rather than "can I change yours", and for a great many projects that is the question that actually matters.

Limitations. We compared what each vendor publishes. We did not test whether an agent in practice overwrites an unlocked hand-edited file, because that varies by prompt and is not something eight vendors document at a comparable level. No plan prices appear on this page, because prices move and plan names do not.