Claude Code to production: a checklist for apps built with Claude Code

To take a Claude Code app to production, treat the agent's output like a new developer's work. Review the git history, prove the tests check real behaviour, hunt for secrets in .claude/settings.local.json, set deny rules and hooks, keep production credentials away from the agent, lock down the database, and plan rollback and monitoring.

Claude Code can build a working app in days. Getting that app safely in front of paying customers is a different job. The agent writes code, runs commands and reports back, and it will often tell you a task is done. Whether the code supports that claim is something you have to check.

This checklist is for founders and teams whose app was largely written by Claude Code and is about to go live. It is not a setup guide: our Claude Code for business guide covers plans, pricing and team rollout. Here we start from an app that already exists. Claude Code features were checked against Anthropic's documentation on 20 November 2026.

  • Review what the agent wrote: git history first, then the largest diffs, then anything touching sign-in, payments or data access.
  • Run the tests yourself. A passing summary from the agent is not proof; failure paths and other users' data need their own tests.
  • Check .claude/settings.local.json for secrets. Apiiro found 1,379 distinct secrets in public copies of the file.
  • CLAUDE.md is guidance only. Use deny rules, hooks and managed settings for anything that must happen every time.
  • Keep production credentials and live customer data out of the agent's reach.
  • Launch with a tested rollback and alerts, not just a deploy button.
Key takeaways

Is Claude Code safe for building a production app?

The tool can be used safely; the output still needs checking. Claude Code has layered controls, from permission modes to hooks, but none of them proves that the code it wrote is correct or secure. Developers know this. In the Stack Overflow Developer Survey 2025, more developers distrusted the accuracy of AI output than trusted it.

How developers rate AI coding tools
  • 84% of developers use or plan to use AI tools in their development process
  • 46% distrust the accuracy of AI tool output, against 33% who trust it
  • 66% say their biggest frustration is AI solutions that are almost right, but not quite
  • 45.2% say debugging AI-generated code is more time-consuming

Source: Stack Overflow Developer Survey 2025, AI section

"Almost right" is the core risk for a production app. The happy path works, the demo looks finished, and the gaps sit in places nobody clicked. Independent scans of AI-built apps show what those gaps look like once they are live.

What a scan of public vibe-coded apps found
  • 5,600+ publicly available apps analysed
  • 2,000+ vulnerabilities found
  • 400+ exposed secrets
  • 175 exposures of personal data, including medical records and IBANs

Source: Escape, Methodology: how we discovered vulnerabilities in apps built with vibe coding, 29 October 2025

Escape's scan covered app builders such as Lovable, Base44 and Bolt.new, not Claude Code specifically, and its authors call the figures a conservative baseline. The lesson carries over: vibe-coded apps fail in the same predictable places, whichever tool wrote them.

How do you review code that Claude Code wrote?

Start with the git history, not the screens. You want to know what changed, when, and which changes nobody read. Three passes cover most apps:

  • History. Read the commit log end to end. If commits list Claude as co-author, filter on that line to see every agent-written change. Look for commits that touch dozens of files with a one-line message.
  • Large diffs. Sort changes by size and read the biggest ones properly. Large agent diffs often hide duplicated logic, dead code and quietly changed behaviour.
  • High-risk paths. Read every line that handles sign-in, roles, payments, file uploads and database queries, whatever its size.

Claude Code can help with the reading. A subagent runs in its own context window with its own allowed tools. Anthropic's example code-reviewer subagent is limited to Read, Glob and Grep, so it can review but cannot edit. Give it your checklist and it returns a consistent second look. It is still a second look, not a sign-off: a person who understands the business should approve what ships.

Do your tests prove the app actually works?

Often not. Agents tend to write tests that confirm the code they just wrote, and a summary saying "all tests pass" can be true while the important behaviour is untested. Check three things yourself.

  • Run them. Run the full suite on a clean checkout, not in the agent's session. Make sure tests were not skipped, deleted or loosened to get a green result.
  • Test real behaviour. Tests should exercise the app the way a user does: sign up, pay, upload, cancel. Mocks that replace the database or payment provider prove very little on their own.
  • Test failure paths. Expired sessions, failed payments, bad input and, above all, one user trying to reach another user's data.

Then make tests part of how Claude Code works. A Stop hook runs when Claude finishes responding. Anthropic's hooks reference says exit code 2 on Stop "prevents Claude from stopping", so a hook that runs your test suite and exits 2 on failure keeps the agent working until the tests pass.

Has Claude Code leaked secrets into your repository?

Check before you assume not. When you choose "Yes, and don't ask again" on a permission prompt, Claude Code saves that approval as an allow rule in .claude/settings.local.json. If the approved command contained a password or token, the secret is saved too, in plain text.

Anthropic's settings documentation says that the first time Claude Code writes this file in a git repository, it adds it to your global git excludes. If you create the file by hand, you must add it to .gitignore yourself. As The Hacker News reported on 7 November 2026, that protection applies only on one computer, does not cover files already committed, and does not travel with the project. Apiiro, the security firm behind the research, found 104,185 public copies of the file on GitHub, of which 1,122 held at least one secret.

Secrets found in public copies of .claude/settings.local.json, by type

1,379 distinct secrets in 1,122 files; selected types shown

Secret typeCount
Database passwords379
SSH logins for public servers43
GitHub, GitLab and npm tokens42
AWS keys23

Source: The Hacker News, Claude Code's 'Don't Ask Again' file exposes passwords and keys, 7 November 2026

Popular secret scanners missed some of these, mostly passwords typed inside commands. So check by hand as well as by tool:

  • Search the repository and its full history for .claude/settings.local.json and .claude/settings.json, and read any rule that contains a credential.
  • Rotate every credential the files have ever held. Deleting the file does not un-leak a secret.
  • Untrack the local file with git rm --cached and add it to the project's own .gitignore.
  • Keep secrets out of commands by reading them from environment variables or a secrets manager.
  • Turn on GitHub secret scanning. It runs automatically for free on public repositories and scans the entire git history; GitHub's advice on a leak is to rotate the credential immediately.

Which permission mode and rules should you use before launch?

Match the mode to the risk. Since Claude Code v2.1.283, interactive terminal and VS Code sessions start in auto mode, where a classifier reviews actions instead of you. Anthropic's permission modes guide is direct: "Auto mode reduces permission prompts but does not guarantee safety."

Claude Code permission modes and what runs without asking
ModeEdits files without askingRuns commands without askingBest fit near launch
default (manual)NoPartial, read-only commandsSensitive code and unfamiliar repos
acceptEditsYesPartial, common file commandsIterating on code you review
planNo, until a plan is approvedPartial, read-onlyInvestigating before a fix
autoYes, classifier-checkedYes, classifier-checkedRoutine work you trust
dontAskPartial, pre-approved onlyPartial, pre-approved onlyLocked-down CI and scripts
bypassPermissionsYesYesIsolated containers and VMs only

Source: Anthropic, Claude Code docs: Permission modes, checked November 2026

Whatever the mode, write allow, ask and deny rules in the shared .claude/settings.json. Anthropic's permissions guide says rules "are evaluated in order: deny, then ask, then allow", so a deny always wins. Deny reads of .env files and secrets folders, and ask before deploys, migrations and anything that deletes data.

Know the limit. Read deny rules cover Claude's file tools and commands it recognises, such as cat, but not a script that opens files itself. For operating-system enforcement, turn on the sandbox. To stop bypass or auto mode being used at all, set permissions.disableBypassPermissionsMode or permissions.disableAutoMode to "disable", ideally in managed settings that nobody can override.

How do hooks and CLAUDE.md keep the agent on track?

CLAUDE.md tells Claude how your project works; hooks make sure certain things happen regardless. Anthropic's memory documentation says Claude treats CLAUDE.md "as context, not enforced configuration" and suggests keeping each file under 200 lines. Put the test command, folder structure, conventions and what not to touch there.

Long sessions add another risk: compaction. When the context window fills, Claude Code summarises the conversation. The project-root CLAUDE.md is re-read from disk afterwards, but an instruction given only in chat can be lost. If a rule matters, move it into CLAUDE.md, and if it must never be broken, enforce it with a hook or deny rule.

ControlWhat it doesEnforced?Use it for
CLAUDE.mdInstructions loaded every session, re-read after compactionNo, guidance onlyTest commands, conventions, no-go areas
Deny rulesBlock matching tool calls before they runYes, for Claude's tools.env files, secrets, destructive commands
PreToolUse hookRuns before a tool call; exit code 2 blocks itYesBlocking production commands and risky edits
Stop hookRuns when Claude finishes; exit code 2 keeps it workingYesRunning tests before Claude says it is done
Reviewer subagentSeparate context with read-only toolsNo, advisoryA consistent second look at changes
Managed settingsOrganisation policy above all other filesYesDisabling bypass and auto mode
Which Claude Code control to use for which job

How do you keep production credentials and data away from the agent?

Give the agent a development environment and nothing else. Unless you turn on its sandbox, Claude Code's commands run with your user's access, so anything your terminal can reach, the agent can potentially reach too.

  • Separate environments. Use separate development and production databases, keys and hosting accounts. Production secrets live in your hosting platform, not in the repository or on the machine where the agent runs.
  • No live customer data. Use seed or anonymised data in development.
  • Database rules. Broken access control is A01 in the OWASP Top 10:2025. On Supabase, Row Level Security must be on for every table in an exposed schema, and the service_role key, which bypasses it, must never reach the browser.

Our vibe-coded app security checklist walks through database policies, keys, storage and payments in detail.

What should deployment, rollback and monitoring look like?

Boring and repeatable. Developers are cautious here: in the Stack Overflow survey, 75.8% said they do not plan to use AI for deployment and monitoring. Before launch, make sure you have:

  • A pipeline. Deploys run from the main branch through automated checks, not from a laptop or an agent session.
  • Rollback. You can return to the previous version in minutes, and you have tried it once.
  • Safe migrations. Database changes are reviewed, backed up first and, where possible, reversible.
  • Monitoring. Error tracking, uptime checks and alerts that reach a person. Security logging and alerting failures are A09 in the OWASP Top 10:2025.

What order should you work through the checklist?

Freeze features first, then work from understanding the code outwards to running it in public.

  1. Freeze new features and tag the current version
  2. Review the git history and the largest diffs
  3. Run the tests yourself and add tests for failure paths
  4. Search the repository and its history for secrets, then rotate them
  5. Set permission rules, hooks and a short CLAUDE.md
  6. Separate production credentials and check database access rules
  7. Set up the deploy pipeline, rollback and monitoring
  8. Launch, then re-check after the first week of changes
A Claude Code pre-launch checklist in eight steps

If steps 2 to 4 turn up a lot, step back before patching. Our guide to whether to fix or rebuild a vibe-coded app helps you decide, and what it costs to fix a vibe-coded app covers budgets. Using OpenAI's agent instead? See the Codex app to production checklist.

When should you get help taking a Claude Code app live?

Before launch, if the app takes payments, stores personal data beyond names and emails, or faces investor or customer due diligence. Also if nobody on the team can confidently explain how sign-in and data access work.

We use these tools ourselves. The Study Room's Portal, Dentorb and ORBQ were built with AI-assisted development, using Antigravity, Claude Code, Codex and Lovable. Circlemount is not an Anthropic partner; we are a web, software and AI agency based in Telford, working with teams across Shropshire, Birmingham and the rest of the UK. Our vibe-coded app security audit works through this checklist with you, and our Claude Code training sets up permissions, hooks and CLAUDE.md for teams who want to keep building.

Frequently asked questions

Is Claude Code safe to use for a production app?

Claude Code itself has strong controls, including permission modes, deny rules, hooks and sandboxing, but none of them proves the code it wrote is correct or secure. Treat its output like a new developer's work: review the diffs, run the tests yourself, check for leaked secrets and test database access before launch.

What is .claude/settings.local.json and why does it matter?

It is Claude Code's personal settings file for one project. When you choose Yes, and don't ask again, Claude Code saves the approved command there, including any password in it. It is excluded from git on your machine only, so check whether it was ever committed and rotate any secrets it held.

Which permission mode should I use on a production codebase?

Use default (manual) or plan mode for sensitive work, such as authentication, payments and database migrations. Auto mode reduces prompts but Anthropic says it does not guarantee safety. Keep bypassPermissions for isolated containers or VMs only, and add deny rules for secrets and destructive commands whichever mode you use.

Does CLAUDE.md stop Claude Code doing risky things?

No. Anthropic says Claude treats CLAUDE.md as context, not enforced configuration. Use it to explain how the project works, how to run the tests and what to avoid. For anything that must never happen, such as reading .env files or running production commands, use deny rules or a PreToolUse hook instead.

What are the most common Claude Code mistakes before launch?

The common ones are trusting the agent's summary instead of running the tests, committing secrets in settings or command history, leaving database access rules too loose, letting the agent reach production credentials, and launching without a rollback plan or alerts. Most are quick to check once you know where to look.

Can Claude Code lose instructions in a long session?

Yes. When the context window fills, Claude Code compacts the conversation into a summary. The project-root CLAUDE.md is re-read from disk, but instructions given only in chat may be lost. Put lasting rules in CLAUDE.md and enforce critical ones with hooks or permission rules, which do not depend on memory.

Sources

Checked on 20 November 2026. Product features and figures change quickly, so check the current source before you rely on a detail.

Getting a Claude Code app ready to launch?

We fix Claude Code apps and get them production-ready: code review, real tests, secret clean-up, database security, deployment and monitoring. The initial app assessment is free, project work starts from £1,000 + VAT, and ongoing support plans start from £125 a month + VAT. Join the Circle for a free business analysis and proposal.