Repos that make vibe coding less painful
A growing list of free, open-source tools that fix the usual problems of building with an AI agent. Each one gets an honest verdict, built from what the people who installed it are saying.
11
Playwright MCP
Lets your coding agent open a real browser and click, type and fill forms in your app, reading the page as structured text instead of screenshots.
Worth a tryMicrosoft's official, free way to let your agent drive a browser through your app's user flows, including in Firefox and Safari's engine. It's token-hungry, and Microsoft's own README now points coding agents to the lighter Playwright CLI, so try both. For debugging speed, network and console problems, pick Chrome DevTools MCP instead.
12
OpenShell
NVIDIA's locked room for AI agents. It runs your agent in a sandbox and only lets it touch the files, websites and keys you've approved.
Only if you need itSerious, well-run security software from NVIDIA, built for teams running many agents at once. For a solo vibe coder it's a lot of setup (Docker, a custom image, API keys, written rules), so start with Claude Code's built-in sandbox and come back if you outgrow it.
13
DeepSeek Harness
DeepSeek's own open-source coding agent, run from a web page or desktop app, built so that almost every feature is a plugin you can swap.
Give it a few weeksA genuinely interesting, very cheap-to-run open-source rival to Claude Code, but it's ten weeks old, labelled a developer preview with breaking changes promised, and every release so far is a pre-release. Check back once it ships a stable version, and switch off its session-log upload if you try it now.
14
OpenCode
The most popular open-source coding agent. It works like Claude Code, in the terminal, a desktop app or the browser, but with almost any AI model you like.
Only if you need itThe most complete open-source alternative to Claude Code, and the one to pick if you want cheaper or open models, or to use your ChatGPT or Copilot plan. It can't use a Claude subscription, its security record this year is poor, and version 2 is still settling, so if Claude Code already works for you, stay put.
15
Spec Kit
GitHub's toolkit for writing down what you want built, as a spec, a plan and a task list, before your coding agent writes any code.
Only if you need itUseful when you want a written record of what your app should do, or you switch between agents or work with other people. For a solo builder on one agent, most users find it heavy, and plan mode or a lighter tool covers most of the same ground.
16
Serena
Gives your coding agent the code-navigation and editing tools your editor has, so it can find, rename and rewrite code by name instead of by searching text.
Only if you need itCan genuinely help on big codebases and big refactors, especially in Codex and other agents. In Claude Code with the newest models you have to fight to make the agent use it, version 2 is days old, and 2026 brought three serious security fixes. Most people building small and medium apps can skip it.
17
shadcn/ui
The buttons, forms, menus and dialogs behind a huge share of React apps, copied into your project as code you own, with a CLI built for coding agents.
Worth a tryIf you build React apps with an agent, this is the safest default for your UI. Agents know it well and the tooling is made for them. Just don't ship the default look, and expect some work when you update components later.
18
Supabase
A hosted Postgres database with login, file storage, an instant API and a dashboard, and the backend most AI app builders reach for first.
Worth a tryA solid default backend, and the one your agent probably already knows best. But your data is only private if every table has row level security with rules that actually restrict access, and AI-written rules often don't. Thousands of apps have leaked data this way. Test it as a stranger and as a second user before launch, and give your agent read-only access.
19
PocketBase
A whole backend in one small program, with a database, user logins, file storage, live updates and an admin dashboard, that you run on your own server.
Only if you need itWonderful for small apps, internal tools and side projects that fit on one cheap server, and far lighter than self-hosting Supabase. But it's still pre-1.0, run by one person, self-hosted only, and its own FAQ tells you not to use it if you'll rely only on an AI tool. Pick it if you're willing to read some docs and look after a server.
20
PostHog
Analytics, session recordings, feature flags and error tracking in one tool, with a big free tier and an MCP server so your agent can set it up and read the numbers for you.
Worth a tryThe best free way to see what users actually do in your app, and its setup wizard and MCP server work well with coding agents. But it records a lot by default, so consent and privacy are your job, data on the US cloud is used to train PostHog's AI unless you opt out, and the bill grows with your traffic. Use the cloud version, set billing limits, and skip self-hosting.
Get the next drop by email.