Six Free Add-Ons That Fix Claude Code's Most Annoying Habits

Picture this: Claude Code produces twenty confident lines built around a library function that doesn't exist and never did. It invented the call, sounded certain, and moved on. Anyone who has used the tool for more than a day knows that moment.

Most of these small irritations already have fixes. The catch is that plenty of users install Claude Code, treat it as a somewhat smarter autocomplete, and never touch what surrounds it: skills, MCP servers and plugins. That ecosystem is where the tool stops being merely helpful and starts removing chores outright.

Six free setups are worth a look. Some come from Anthropic, GitHub and Microsoft, and a couple are community projects. Each section gives the problem it solves, the reason it's useful and the install command, and the final part explains which to try first.

Context7: stop the made-up API calls

The classic complaint is code that looks right but calls functions from a library version two years old, or ones that never shipped. Context7 attacks that directly. It fetches the current, version-specific documentation for whatever library is in play and places it into context. Claude then writes against the library as it exists today rather than as it looked at training time.

Think of it as handing the model a fresh copy of the docs on every request, so it doesn't work from memory. After installation it fades into the background, and the wrong-API problem simply stops showing up.

Why it earns a place: confidently wrong library calls are probably the most irritating habit of AI coding, and this is the cure.

Install it with:

claude mcp add - scope user context7 - npx -y @upstash/context7-mcp - api-key YOUR_API_KEY

A free key from the Context7 dashboard gives better rate limits.

Playwright MCP: let the AI test its own work

Microsoft's Playwright MCP hands Claude an actual browser to use. It can load your app, press buttons, complete forms and verify that features behave. Control runs through the accessibility tree instead of screenshots, so it's fast and pixel noise doesn't throw it off.

A typical flow: build a feature, then say something like "open the app and walk through the signup flow." Claude launches the browser, runs through it and reports where it broke. There are no hand-written test scripts and no window switching.

Why it earns a place: the AI wrote the code, so it can click through the result and check it, giving you QA that needs no babysitting.

Install it with:

claude mcp add playwright npx @playwright/mcp@latest

The docs carry a caution worth repeating. This isn't a security boundary, so don't aim it at anything sensitive and walk away.

GitHub MCP: repo chores in plain language

For anyone on GitHub, which is most people, the official GitHub server connects Claude directly to the platform. Issues, pull requests, code review and CI status all become available through natural language, without leaving the terminal.

Typical requests might read like this:

• Ask for the oldest open issue to be triaged, with a draft fix attached.
• Have a pull request opened for the current branch, complete with a decent description.
• Find out what broke in the most recent CI run.

Because it's published by GitHub itself, trusting it with access to your repositories is an easier call.

Why it earns a place: routine admin such as triage, PR descriptions and chasing CI turns into something you ask for instead of something you do.

Install it with:

claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'

A personal access token with suitable scopes is required.

Anthropic's document skills: real files, not scripts

This is Anthropic's own pack, and it's a gentle starting point for anyone who has never installed a skill. The document skills are the piece to reach for. Normally, asking for a report gets you text, or a script that would generate a document. With the skills installed, Claude produces the actual file: a proper PDF, a formatted DOCX, or an XLSX spreadsheet with the numbers sitting in cells.

The pack doubles as a tidy example of a well-built skill, handy for anyone planning to write one.

Why it earns a place: asking for a spreadsheet should return a spreadsheet, not Python code that makes one.

Install it with:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills

Spec Kit: agree on the spec before any code

GitHub's Spec Kit targets a familiar failure. A vague prompt goes in, Claude quickly builds a large amount of the wrong thing, and the next hour goes to untangling it.

Spec Kit asks for a short spec first, covering what is wanted and why, before code gets written. Claude then works from that spec instead of guessing. It sounds like extra effort, but it saves effort, because less wrong output gets thrown away. Tiny tasks don't need it. For anything with real scope, settling the spec first can save more time than any other item here.

Why it earns a place: the bulk of wasted AI-coding time is Claude building the wrong thing fast, and a spec up front prevents it.

Install it with:

uv tool install specify-cli
specify init my-project --ai claude

Python 3.11 or newer and the uv package manager are prerequisites.

Sentry MCP: debug production from the real crash

The last one is for when something is already broken in production. Sentry MCP pulls actual errors and stack traces into Claude. Rather than pasting a trace out of a dashboard, you can ask it to look at the error that spiked an hour ago and work out the cause. Claude reads the live Sentry data and works from the genuine crash, not a description of it.

That closes the loop: an error occurs, you ask, and a fix comes back, all grounded in the true stack trace.

Why it earns a place: production debugging usually means five browser tabs and plenty of copy-paste, and this puts the error where Claude can read it.

Install it with:

claude plugin marketplace add getsentry/sentry-mcp
claude plugin install sentry-mcp@sentry-mcp

A Sentry account and an auth token with read access are needed.

A note on the three kinds of add-ons

The six picks span three formats. Context7, Playwright, GitHub and Sentry are MCP servers, meaning external programs Claude talks to through a standard protocol. The document skills are a skill pack, a bundle of instructions and helper files Claude loads when a task calls for them. Spec Kit is a command-line tool that scaffolds a spec-first workflow in a project. Knowing which format a tool uses explains why the install commands look so different from one another.

Install all six at once?

No. Installing everything on day one is a good route to feeling overwhelmed and using none of it. A better approach is to pick one, wire it into something under way this week, and let it prove itself before adding the next.

Anyone who writes or generates documents a lot should start with the document skills, which takes about a minute. If wrong API calls are the main annoyance, Context7 alone will improve the week.

The real skill isn't collecting tools. It's spotting which dull part of the day a tool quietly removed, then looking for the next one.

Quick answers

Which should go in first? Context7 for anyone tired of hallucinated APIs, or the document skills for the fastest pleasant surprise. Either takes a minute.

Are they safe? The official ones from Anthropic, GitHub, Microsoft and Sentry are about as safe as these things get. For anything that touches files, repositories or a browser, read what access it requests before granting it. A tool able to act on your behalf deserves ten seconds of reading.

Do any of them cost money? No, they all install free. A few connect to services that may already carry a bill, such as Sentry and GitHub, but the tools themselves don't charge anything.