By the time you close this page you will have a short install list and a clear reason to skip the rest. Most people who wire up Claude add-ons spend an afternoon and get one tool working plus three broken configs.
The lists they work from rank by novelty, so the newest server sits at the top. The ranking that helps is by catch. MCP, short for Model Context Protocol, is the standard that lets Claude reach outside tools like GitHub, a database or your meeting notes through a small connector called a server.
These nine are ordered by the limit you can live with, not the length of the feature list. For each one you get the problem it fixes, the catch that rules it in or out, and the first setup move. The catch comes before the payoff every time.
Pick two or three that match the work you already do this afternoon, and close the tab on the rest.
| Add-on | What it fixes | Main catch | Where it runs |
|---|---|---|---|
| Context7 | Hallucinated library functions and stale APIs | Only covers the libraries it indexes | Claude Code and MCP clients |
| Superpowers | Guessing instead of planning and testing | Strict red-green-refactor fights quick scripts | Claude Code plugins |
| GitHub MCP | Copy-pasting issues, PRs and Actions logs | Large tool surface crowds context | Claude with GitHub repo access |
| Playwright MCP | Testing from source instead of a real page | Slower and heavier than API calls | Claude Code |
| Sentry MCP | Debugging imaginary incidents instead of real ones | Human-in-the-loop, not full API replacement | Claude with Sentry project access |
| Linear MCP | Retyping ticket summaries after decisions | One OAuth session per workspace | Claude with Linear workspace |
| Supabase MCP | Guessing schema and risky queries | Built for dev and test, not production | Claude with dev database |
| Granola MCP | Exporting meeting notes before Claude can read them | Basic limits notes to last 30 days; paid for transcripts | Claude with Granola account |
| Claude for Google Workspace | Switching tabs to draft in a separate chat | Cannot handle comments or suggestions in Docs | Google Docs sidebar (beta) |
1. Context7 for Version-Pinned Docs
The way to stop Claude inventing library functions is to hand it the real documentation for the exact version you installed. Context7 does that, serving version-pinned docs on demand.
It is hosted at https://mcp.context7.com/mcp, under an MIT license, and needs no API key. That's the credential string a service usually asks you to paste in to prove who you are.
The failure it kills is quiet. Claude calls a function renamed two versions ago, or passes an argument that never existed. The code looks right on screen and fails the moment it runs.
Prompt to copy
Here is a call Claude wrote to our database library: db.findMany({ where: { active: true }, take: 25 }). Our project is pinned to an older major version of that library. Check this call against the documentation for that exact version, quote the signature you find, and rewrite the call if it is wrong. Give me the corrected line only, plus the version you checked.
That's the API, the set of functions a library exposes for code to call. Where a library sits inside the index, the answer names the right signature in seconds instead of guessing.
The catch: Context7 only covers libraries it has indexed. Outside that set, pair it with web search. The fix is bounded by the index, not by the model.
For anyone who codes with Claude, this is the highest-impact fix for invented function calls, and it earns the first slot. Your first weekend move is adding that endpoint as a server, with nothing else attached.
💡 Tip: Add web search next to Context7 for libraries it does not index. The pair covers most of what you will hit.
The Add-On Vocabulary, Decoded
2. Superpowers Plugin for Disciplined Workflow
Tests fail every run. The plugin enforces red-green-refactor, where a test must fail before any line of implementation is written.
That makes it a poor fit for a one-off script you want done in five minutes. The safe pattern is to write the test first and accept the extra step on code you plan to keep.
Jesse Vincent's community plugin arrived in one command, and the docs, version pins and working signatures came with it.
Superpowers adds a second layer by bundling more than a dozen skills in one install: brainstorming, planning, systematic debugging and test-driven development among them.
Install is a single line typed into Claude Code.
Prompt to copy
I'm adding a small feature to my project. Work through it in this order. 1) Ask me what the feature should do. 2) Write a failing test and show me the red output. 3) Write the smallest change that passes. 4) Rerun the suite and show me the green. Do not skip the failing step, and do not change my existing tests.
A small password reset took about 40 minutes by hand across test files and reruns, and the plugin's own answer came back in under a minute.
Reviewing the plan and the failing test took about 15 minutes, and only then did the fix pass. Say 55 minutes against the hour, for a codebase you will keep.
3. GitHub MCP for Repository Work
GitHub's official hosted MCP server lets Claude read and work with your repository directly. It covers issues, pull requests, code search and Actions logs, and it is free with any GitHub plan.
A disciplined workflow leaves branches, commits and pull requests behind, and this is the add-on that puts those in front of Claude.
Connect: add GitHub's hosted server to Claude's MCP settings and authorize it against your account. You should see the connection listed as active before you type anything.
Limit the toolsets: at setup, switch on only the groups you need, such as pull requests and Actions. You should see a much shorter tool list.
That matters because every tool eats part of the context window. That's the fixed amount of conversation Claude can hold at once, and a crowded one degrades every answer.
Ask: "Find the open pull request that touches the login page, read the failing Actions log on it, and tell me in three lines what broke." The answer should name the PR and quote the failing step.
Catch: never paste passwords, client files, or medical or financial records into the chat, or into the issues, pull requests and logs you connect it to.
The payoff is a PR found, a failing log read and code searched, with no copy-paste and no browser tabs open.
💡 Tip: Start with issues and pull requests only. Add Actions logs once you have confirmed the basics work.
4. Playwright MCP for Real Browser Checks
A signup form that looks right in the code and fails in the browser is the tedious part.
You click through fields, submit, watch the page, and repeat it after every change. Microsoft's Playwright MCP server hands Claude Code a real browser to do that pass for you.
It loads the page, fills the form and screenshots the result, so the check runs against what renders, not what the source says it should. That's the front end: the part of the page a visitor actually sees and clicks.
The catch is weight. Playwright is slower and heavier than an API call, because it drives a full browser to get its answer. When the job is only reading a page's text, web_fetch is the cheaper tool, and it finishes faster.
Save Playwright for jobs where the outcome depends on rendering. A form that has to submit without an error. A layout that has to look right at phone width. For everything that is just words on a page, keep web_fetch and skip the browser.
Playwright vs web_fetch — When to Reach for Which
📐 Watch for
Widths beyond desktop: 375px phone, 768px tablet
🚨 If the form fails
Screenshot shows the error banner – report that, not the source
🔑 Skip it when
Page needs a login, a paid plan or an API key – use web_fetch
⏱️ Check the run took
Longer than web_fetch – budget for it before a big test pass
Fast test
Before wiring it, ask: would the same page read fine as text in web_fetch? If yes, do not open a browser.
5. Sentry MCP for Production Debugging
An error that only shows up for real users eats an afternoon. You have a stack trace, no reproduction and a guess about which release introduced it.
Sentry's hosted MCP server brings the production errors, traces and release health into the same conversation where Claude is debugging.
The catch is scope. Sentry built this for human-in-the-loop debugging, where the tool guides the investigation and a person makes the call. It is not a full replacement for the Sentry API, so expect guided digging, not bulk automation across every setting.
A typical session runs about 20 minutes of Claude reading traces, 6 minutes of your review of the proposed fix, and no extra cost beyond your existing Sentry plan. That is a fraction of an afternoon spent chasing a stack trace by hand.
⚠️ COMMON MISTAKE
Do not paste the client file to save a step
It is tempting to drop a spreadsheet with customer records into the chat so Claude can spot the pattern. Do not. Paste a header row and one anonymized sample instead.
6. Linear MCP for Issue Tracking
Prompt to copy
Read the last three messages in this chat and write one Linear issue for the follow-up work we agreed on. Title it in six words or fewer, put the open question in the description, set priority to Medium and assign it to me. Do not create it. Show me the title and description first so I can correct them.
That runs from inside Claude, through Linear's hosted connector, so the ticket gets written without a second tab.
The server finds, creates and updates issues, projects and comments, and it authenticates over an encrypted connection with a sign-in token rather than a pasted key.
Setup takes about ten minutes, and the answer comes back in seconds.
The catch bites anyone in more than one workspace. Each sign-in session is tied to a single workspace, so the first switch lands your issue in the client's board or another team's backlog.
Check the workspace name in Claude's reply before you confirm, and sign in separately for the second one.
Worth it if chat keeps producing follow-up work that never reaches a tracker. A decision, a small bug, a status change becomes a real issue you can assign and close, with no retyping of the summary you just settled.
7. Supabase MCP for Safe Database Work
A wrong draft query against a test database costs you a rollback, not an incident. That safety comes from one install, and the order of two moves. Tickets eventually point at data, and the database is the next surface to touch.
Supabase built its MCP server for development and testing and warns against connecting it to production data. That warning is the install condition, not small print. Add read_only=true before the first query, never after.
Then Claude reads the schema and hands you queries to run, while the read-only user caps any slip. Setup takes one sitting; every later session starts safe.
💡 Tip: Point it at a dev or test database only. If you cannot name the environment, do not connect.
Supabase MCP: two ways to run it
Read-only user
- Runs every query as a read-only Postgres user
- Wrong drafts cannot delete or overwrite
- Connect a dev or test database
- Safe for schema browsing and query drafting
Full access
- Can change data and structure
- Production data is explicitly warned against
- A bad draft can delete rows
- Reserve for controlled local environments
8. Granola Connector for Meeting Notes
Yes, and without a single export. Granola's official MCP connector sets up in under 5 minutes and syncs your meeting notes straight into Claude, so the decisions behind your schemas arrive without a copied file.
The old way: find last Tuesday's call, download the notes, paste them in, lose track of which quote came from where. The new way: ask Claude what was decided about the schema, and every answer links back to the transcript with a double-click.
The catch: the free Basic plan reaches only your personal notes from the last 30 days. Transcript access needs a paid plan, and on Enterprise an admin has to switch MCP on first.
On a paid plan that covers your recent meetings, decisions and action items come out of the chat, each claim one click from the original recording.
9. Claude for Google Workspace in Docs
Claude now sits in the Docs sidebar and rewrites a paragraph without you switching tabs. The Workspace beta is open on Pro, Max, Ultra, Team and Enterprise plans, and the document is the last surface in this tour, right where meetings end up.
Where not to hand it over: anything collaborative. From the Docs sidebar Claude cannot read, reply to or resolve comments. It also cannot accept or reject existing suggestions.
So the split is fixed by the beta itself. Let Claude draft and rewrite inside the document while a person handles the comment thread and the tracked changes, which is where judgment and authorship live anyway.
Used that way, the beta saves you the copy-paste round trip and leaves the reviewing to the people who should own it.
🧭 Install in the order that matches your work
Start with Context7
Install this first if you write any code with Claude, because stale library calls waste the most time.
Add Superpowers next for planned work
Install it when your week includes features with tests, not just one-off scripts.
Sentry for production debugging
Add this when you already respond to real incidents and have a Sentry project to connect.
Linear for ticket work
Add this when most of your day is creating and updating issues, and authorize one workspace at a time.
Granola for meetings
Add this if you sit in client calls and need to pull decisions out of recent transcripts.
Workspace for document drafting
Add this last if you already draft in Google Docs and can live with its beta limits.
Frequently Asked Questions
Do I need to install all nine?
No. Two or three cover most people. Start with Context7 if you code, then add the one that matches your daily work, whether that is tickets, production errors, meetings or document drafting.
Are these add-ons free?
Some are. Context7 is MIT-licensed and needs no API key. GitHub MCP is free with any GitHub plan. Others depend on your plan: Granola's transcript access needs a paid plan, and Claude for Google Workspace is a beta for Pro, Max, Ultra, Team and Enterprise.
What is an MCP server, in plain words?
It is a small service Claude connects to so it can call that service's tools during a conversation. Once connected, Claude can read a ticket or fetch documentation without you copying and pasting.
Is it safe to connect Claude to my GitHub or database?
Yes, with limits. Never paste passwords or client files into the chat. For Supabase, connect a dev or test database and set read_only=true. For GitHub, keep the connection to the repos and toolsets you actually need.
Do I need to know how to code to use these?
Some of them, yes. Context7, GitHub MCP, Playwright and Supabase assume you work with code. Granola and Claude for Google Workspace are the two that fit a non-programmer's day most easily.
Which two to install this afternoon
Context7 first if you write any code with Claude. It fixes the problem that wastes the most time, and the install is one endpoint with no API key. Superpowers second if your work includes planned features with tests.
Then pick the one that matches your day. Sentry for production debugging. Linear for ticket work. Granola for meetings. Workspace for Docs drafting. Skip the rest until the problem shows up.
The order matters less than the fit. Install the ones you will use this week and leave the others for when the catch no longer applies.