Category: Guides & Manuals

Last Verified: September 2026 — commands checked against exa-labs/exa-mcp-server and the official Claude Code MCP docs

Ask Claude Code about a library that shipped last month and one of two things happens: it answers from training data that’s already stale, or it searches and comes back with a list of links and snippets it then has to fetch one by one. Either way you spend the next few turns steering it toward the page that actually has the answer.

Exa MCP connects Claude Code to Exa, a search engine built for AI agents. Instead of ten blue links, web_search_exa returns clean page content ready to read, and web_fetch_exa turns any URL into Markdown. It’s a hosted server, so setup is one command, and it works without an API key to start.

Already using our Firecrawl MCP guide? They pair well: Exa is best at finding the right pages, Firecrawl at scraping JavaScript-heavy sites once you know where to look.

claude mcp add --transport http --scope user exa https://mcp.exa.ai/mcp

You have shell access and the `claude mcp` CLI on this machine. Connect the Exa MCP server to Claude Code and report back. Do not tell me it's done unless step 4 actually confirms it:

1. Run `claude mcp list`. If an exa entry already exists, show me its URL and stop instead of adding a second one.
2. Add Exa's hosted server for all projects (no key needed to start):
   `claude mcp add --transport http --scope user exa https://mcp.exa.ai/mcp`
3. Run `claude mcp list` and confirm exa shows Connected.
4. Use `web_search_exa` to find the latest release notes for Claude Code, then use `web_fetch_exa` on the top result and give me 3 bullet points from the page with its URL.

If I later want higher limits, tell me to create a key at https://dashboard.exa.ai/api-keys and re-add the server with an `x-api-key` header. Never ask me to paste the key into this chat, and never put it in the URL. If any step fails, give me the exact error and the command that produced it. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

claude mcp add --transport http --scope user exa https://mcp.exa.ai/mcp

claude plugin install exa@claude-plugins-official

claude mcp add --transport http --scope user exa https://mcp.exa.ai/mcp \
  --header "x-api-key: <your-exa-api-key>"

claude mcp add --transport http --scope user exa \
  "https://mcp.exa.ai/mcp?tools=web_search_exa,web_fetch_exa,web_search_advanced_exa"

What Each Piece Does

Verify It’s Actually Working

claude mcp list

Then give Claude a question its training data can’t answer:

Use Exa to find what changed in the latest release of the framework this repo uses, and tell me if any of it affects our code

If Claude calls web_search_exa, reads a result, and cites a URL from the last few weeks, it’s working.

Common Mistakes to Avoid

• Adding only the advanced tool. ?tools=web_search_advanced_exa replaces the defaults, so basic search and fetch disappear. List every tool you want in one comma-separated value.
• Leaving the URL unquoted. A ? or & in the URL can be eaten by your shell. Wrap the URL in quotes whenever it has parameters.
• Putting the API key in the URL. It works, but the key ends up in your Claude Code config and shell history in plain sight. Use the x-api-key header.
• Registering it twice. Installing the plugin and running claude mcp add gives Claude two copies of every Exa tool. Pick one.
• Trusting search results blindly. Web pages can contain prompt-injection text. Ask Claude to cite its sources, and review anything it plans to run based on what it read.

Q&A

Official Resources

• exa-labs/exa-mcp-server on GitHub — tools, parameters, and auth options
• Exa MCP documentation — official reference
• Exa API keys — create a key for higher limits
• Claude Code MCP docs — claude mcp add options and headers

Last Verified: September 2026 — commands checked against grafana/mcp-grafana (v1.5.1) and the official Claude Code MCP docs

When something breaks in production, the usual Claude Code loop is: open Grafana, find the right dashboard, squint at a spike, copy a PromQL query or a chunk of Loki logs, paste it into the terminal, and repeat when Claude asks for a wider time range. You end up as a slow, lossy copy-paste bridge between your metrics and your agent.

Grafana MCP is Grafana Labs’ official MCP server. It lets Claude Code search your dashboards, run PromQL against Prometheus and LogQL against Loki, read alert rules and their firing state, and look up incidents on its own. It works with self-hosted Grafana 9.0+ and Grafana Cloud, and it’s open source (Apache 2.0).

Pair it with our Sentry MCP guide: Sentry tells Claude which error is happening, Grafana shows when it started and what the metrics were doing around it.

claude mcp add --scope user --transport stdio \
  --env GRAFANA_URL=http://localhost:3000 \
  --env GRAFANA_SERVICE_ACCOUNT_TOKEN=<your-token> \
  grafana -- uvx mcp-grafana --disable-write

You have shell access and the `claude mcp` CLI on this machine. Connect the Grafana MCP server (grafana/mcp-grafana) to Claude Code in read-only mode and report back. Do not tell me it's done unless step 5 actually confirms it:

1. Run `uvx --version`. If uv is missing, install it with the official installer for this OS (macOS/Linux: `curl -LsSf https://astral.sh/uv/install.sh | sh`, Windows PowerShell: `powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"`), then re-check.
2. Run `claude mcp list`. If a grafana entry already exists, stop and show me its config instead of adding a second one.
3. Ask me for two values: my Grafana URL (e.g. http://localhost:3000 or https://myinstance.grafana.net) and a service account token. Tell me to create the token in Grafana under Administration > Users and access > Service accounts, with the Viewer role. Never echo the token back in chat or write it to any file other than the Claude Code config.
4. Add the server for all projects, read-only:
   `claude mcp add --scope user --transport stdio --env GRAFANA_URL=<url> --env GRAFANA_SERVICE_ACCOUNT_TOKEN=<token> grafana -- uvx mcp-grafana --disable-write`
5. Run `claude mcp list` and confirm grafana shows Connected. Then call the Grafana search tool and list the first 5 dashboard titles.

If any step fails, give me the exact error and the command that produced it. A 400 "id is invalid" error on datasource tools means Grafana is older than 9.0, so say that instead of retrying. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

uvx --version

docker run -d --name grafana -p 3000:3000 grafana/grafana

claude mcp add --scope user --transport stdio \
  --env GRAFANA_URL=http://localhost:3000 \
  --env GRAFANA_SERVICE_ACCOUNT_TOKEN=<your-token> \
  grafana -- uvx mcp-grafana --disable-write

What Each Piece Does

Verify It’s Actually Working

claude mcp list

Then give Claude a real question:

Search my Grafana dashboards, pick the one for the API service, and tell me if the error rate changed in the last 6 hours

If Claude calls a dashboard search tool, then a Prometheus or Loki query tool, and answers with actual numbers from your data, the token, the URL, and the datasource access are all working.

Common Mistakes to Avoid

• Giving the token Admin because it’s easier. An AI agent with an Admin token can delete dashboards and alert rules. Start with Viewer and --disable-write, and move up only when you need a write tool.
• Putting server flags before the --. --disable-write belongs to mcp-grafana, not to claude mcp add. Placed before the double dash, Claude Code rejects it or ignores it.
• Using the old GRAFANA_API_KEY variable. It still works but is deprecated. Use GRAFANA_SERVICE_ACCOUNT_TOKEN for new setups.
• Asking for a whole dashboard. Full dashboard JSON can eat a large chunk of the context window. Ask for the summary or specific panels instead; the server has dedicated tools for exactly that.
• Running Grafana older than 9.0. Datasource tools fail with 400 … “id is invalid”. The fix is upgrading Grafana, not reconfiguring the server.

Q&A

Official Resources

• grafana/mcp-grafana on GitHub — README, full tool list, and every CLI flag
• Grafana service accounts — how to create the token and pick a role
• Claude Code MCP docs — claude mcp add options and scopes

Last Verified: September 2026 — commands checked against Notion’s official MCP docs and the Claude Code MCP docs

Your specs, meeting notes, and bug tickets live in Notion. Your code lives in Claude Code. So every task starts with the same ritual: open the Notion page, copy the requirements, paste them into the terminal, and hope you didn’t miss the comment thread at the bottom where the scope changed.

Notion MCP is Notion’s official hosted MCP server at mcp.notion.com. Claude Code signs in with your Notion account through OAuth and can then search your workspace, read pages and databases as Markdown, and create or update pages, all limited to what you can already see. There’s no API token, no JSON config, and nothing to run locally.

Heads-up: search results still point to the older self-hosted package (@notionhq/notion-mcp-server). Notion’s own README now says that repo is no longer actively maintained and recommends the hosted server instead. That’s the version this guide sets up, for the same reason our community MCP picks skip abandoned servers.

claude mcp add --transport http --scope user notion https://mcp.notion.com/mcp

You have shell access and the `claude mcp` CLI on this machine. Connect Notion's official hosted MCP server to Claude Code and report back. Do not tell me it's done unless step 5 actually confirms it:

1. Run `claude mcp list`. If a notion entry already exists, show me its command or URL. If it runs `npx @notionhq/notion-mcp-server` (the old self-hosted server), tell me and ask before replacing it.
2. If there's no hosted entry, add it for all projects:
   `claude mcp add --transport http --scope user notion https://mcp.notion.com/mcp`
3. Tell me to run `/mcp` inside Claude Code (or `claude mcp login notion` in a terminal) and complete the Notion sign-in in my browser. Wait for me to confirm. Never ask me for a Notion token or integration secret; this server uses OAuth only.
4. Run `claude mcp list` and confirm notion shows Connected.
5. Search my Notion workspace for any page (use a short keyword like "meeting") and show me the titles of up to 3 results. Read only; do not create or edit any page.

If any step fails, give me the exact error and the command that produced it. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

claude mcp add --transport http --scope user notion https://mcp.notion.com/mcp

claude mcp login notion

What Each Piece Does

Verify It’s Actually Working

claude mcp list

Then give Claude a real task:

Find the Notion page for our next feature, list the requirements that don't have a matching file in this repo yet, and don't edit anything in Notion

If Claude calls notion-search, then notion-fetch, and compares the page against your code, the connection and your permissions are working.

Common Mistakes to Avoid

• Following an old tutorial. Guides that tell you to create an internal integration, copy an ntn_ token, and share each page with it are for the self-hosted server. The hosted server needs none of that.
• Keeping both servers. If an old notionApi entry is still registered, Claude sees two sets of Notion tools. Remove it with claude mcp remove notionApi.
• Letting Claude edit pages unreviewed. Notion’s security guide warns that page content can carry prompt-injection text. Keep Claude Code’s permission prompts on for Notion write tools and read what it plans to change.
• Hammering search. notion-search is limited to 30 requests per minute. If you hit rate-limit errors, ask Claude to run fewer searches in parallel.
• Signing in to the wrong workspace. The OAuth page lets you pick a workspace. If search returns nothing, run claude mcp logout notion and sign in again with the right one.

Q&A

Official Resources

• Connect Notion MCP — official setup for Claude Code and other clients
• Notion MCP supported tools — every tool, plan limits, and rate limits
• Notion MCP security best practices — prompt injection and admin controls
• Claude Code MCP docs — scopes, OAuth, and claude mcp commands

Last Verified: September 2026 — commands checked against docker/mcp-gateway (v0.43.3) and the official Docker MCP Toolkit docs

Every MCP server you add to Claude Code is another claude mcp add line, another Node or Python process running on your machine with your user’s full permissions, and often another API key sitting in a config file. By the fifth server, it’s hard to say what’s installed, what it can touch, or where the keys are.

Docker MCP Gateway collapses all of that into a single entry. Claude Code connects to one server, MCP_DOCKER, and the gateway runs every MCP server you pick from Docker’s catalog, each in its own container. Containers are capped at 1 CPU and 2 GB of memory and get no host filesystem access unless you grant it. Credentials live in Docker Desktop instead of your config files. It’s open source (MIT) and ships built into Docker Desktop as the MCP Toolkit.

This guide uses GitHub and Playwright as the example servers. If you set up Playwright the direct way with our Playwright MCP guide, this is the containerized alternative.

docker mcp profile create --name dev-tools

docker mcp profile server add dev-tools \
  --server catalog://mcp/docker-mcp-catalog/github-official \
  --server catalog://mcp/docker-mcp-catalog/playwright

docker mcp client connect claude-code --profile dev-tools --global

You have shell access, Docker, and the `claude mcp` CLI on this machine. Set up the Docker MCP Gateway for Claude Code and report back. Do not tell me it's done unless step 6 actually confirms it:

1. Run `docker mcp --help`. If the command is missing or says Docker Desktop isn't running, stop and tell me to start Docker Desktop and turn on Settings > Beta features > Enable Docker MCP Toolkit.
2. Run `claude mcp list`. If MCP_DOCKER already exists, stop and tell me which profile it uses instead of adding a second one. Also list any existing github or playwright entries, since those would duplicate the gateway's tools.
3. Run `docker mcp profile list`. If there's no dev-tools profile, create it: `docker mcp profile create --name dev-tools`
4. Add two servers to it:
   `docker mcp profile server add dev-tools --server catalog://mcp/docker-mcp-catalog/github-official --server catalog://mcp/docker-mcp-catalog/playwright`
5. Connect Claude Code to the profile for all projects: `docker mcp client connect claude-code --profile dev-tools --global`
6. Run `claude mcp list` and confirm MCP_DOCKER shows Connected, then run `docker mcp tools count` and tell me the number.

The GitHub server needs OAuth. Tell me to authorize it in Docker Desktop (MCP Toolkit > the server's Configuration tab > OAuth) instead of asking for a token. Never ask me to paste a token into this chat. If any step fails, give me the exact error and the command that produced it. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

docker mcp --help

docker mcp profile create --name dev-tools

docker mcp catalog server ls mcp/docker-mcp-catalog

docker mcp profile server add dev-tools \
  --server catalog://mcp/docker-mcp-catalog/github-official \
  --server catalog://mcp/docker-mcp-catalog/playwright

docker mcp client connect claude-code --profile dev-tools --global

claude mcp add MCP_DOCKER --scope user -- docker mcp gateway run --profile dev-tools

docker mcp tools count

docker mcp profile tools dev-tools --disable <server>.<tool>

What Each Piece Does

Verify It’s Actually Working

claude mcp list

Then run the test prompt from Docker’s own docs:

Use the GitHub MCP server to show me my open pull requests

If Claude calls a GitHub tool through MCP_DOCKER and lists real pull requests, the gateway, the container, and the OAuth credential are all working.

Common Mistakes to Avoid

• Skipping the Beta toggle. The MCP Toolkit is off by default. Without it, docker mcp commands fail or behave unexpectedly.
• Keeping the old direct entries. If github or playwright is still registered with claude mcp add, Claude sees every tool twice. Remove the direct entries once the gateway works.
• Forgetting --global. Without it, the connection only applies to the git repo you ran it in, and other projects won’t see MCP_DOCKER.
• “Docker Desktop is not running” on WSL2 or Docker CE. Set export DOCKER_MCP_IN_CONTAINER=1, and outside Docker Desktop, run docker mcp feature enable profiles once.
• Loading every server into one profile. Each server adds tools to Claude’s context. Use separate profiles per project and disable tools you never call.

Q&A

Official Resources

• Get started with Docker MCP Toolkit — official setup walkthrough
• Use MCP Toolkit from the CLI — every docker mcp command used here
• docker/mcp-gateway on GitHub — source code and releases
• Docker MCP Catalog — browse available servers

Last Verified: September 2026 — commands checked against the official getsentry/sentry-mcp repo (0.39.0) and mcp.sentry.dev

A production error fires, and the copy-paste loop starts: open Sentry, copy the stack trace, paste it into Claude Code, go back for the breadcrumbs, paste again, then realize you needed the tags too. The agent only ever sees what you remembered to copy.

Sentry MCP cuts you out of that loop. It’s Sentry’s official server, built for coding agents like Claude Code. The agent can search your issues, pull the full stack trace, breadcrumbs, tags, and trace for an event, and ask Seer, Sentry’s AI debugger, for a root cause. It then fixes the code in the same session. It’s hosted with OAuth sign-in, so there’s no token to create and nothing to install.

This pairs well with our Chrome DevTools MCP setup. DevTools MCP catches bugs on your machine, and Sentry MCP brings in the ones your users hit in production.

claude mcp add --transport http --scope user sentry https://mcp.sentry.dev/mcp

You have shell access and the `claude mcp` CLI on this machine. Connect the hosted Sentry MCP server and report back. Do not tell me it's done unless step 4 actually confirms it:

1. Run `claude mcp list`. If a server named sentry already exists, stop and tell me how it's configured instead of adding a second one.
2. Add the hosted server (OAuth, no token or Node.js needed):
   `claude mcp add --transport http --scope user sentry https://mcp.sentry.dev/mcp`
3. Run `claude mcp list` again and confirm sentry is listed. It's fine if it says it needs authentication.
4. Tell me to restart Claude Code and run `/mcp` to sign in to Sentry in my browser. After I confirm, call the whoami tool and show me which Sentry user and organizations it sees.

Never ask me to paste a Sentry token into this chat. If any step fails, give me the exact error and the command that produced it. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

claude mcp add --transport http --scope user sentry https://mcp.sentry.dev/mcp

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/your-org-slug/your-project-slug

/mcp

claude plugin marketplace add getsentry/sentry-mcp

claude plugin install sentry-mcp@sentry-mcp

claude mcp add --scope user --env SENTRY_ACCESS_TOKEN=your-token --env SENTRY_HOST=sentry.example.com sentry -- npx @sentry/mcp-server@latest

What Each Piece Does

Tool counts come from Sentry’s published skill definitions (0.39.0). The ones you’ll use most are search_issues, get_issue_details, get_event_stacktrace, get_issue_breadcrumbs, and analyze_issue_with_seer.

Verify It’s Actually Working

claude mcp list

Then give it a real production bug to work through:

Find the most frequent unresolved error in [YOUR PROJECT] from the last 7 days. Show me the stack trace and breadcrumbs, tell me the root cause, then open the file in this repo where it happens and propose a fix.

If the tool calls show search_issues and get_event_stacktrace, and the answer cites real event data, it’s working end to end.

Common Mistakes to Avoid

• Thinking “Needs authentication” means it’s broken. Every hosted connection uses OAuth. Run /mcp once and sign in. There’s no token to create.
• Granting Triage when you only wanted to read. Triage lets the agent resolve and reassign issues. Leave it unchecked on the approval screen unless you actually want that.
• Wrong slugs in a scoped URL. /mcp/{org}/{project} uses the slugs from your Sentry URL, not the display names. A typo means the connection fails.
• Running both the plugin and a manual server. Pick one. Two entries mean duplicate tools and a subagent that may not be the one answering.
• Expecting AI search on self-hosted with no LLM key. The stdio server leaves search_issues / search_events unavailable until an LLM provider is configured, and Seer may not be available on self-hosted installs at all.

Q&A

Official Resources

• Sentry MCP (mcp.sentry.dev) — official setup page for every client
• getsentry/sentry-mcp on GitHub — source, plugin, self-hosted instructions
• Claude Code plugin docs — how the sentry-mcp subagent works
• Claude Code MCP Documentation — the claude mcp command reference

Last Verified: September 2026 — commands checked against the official Firecrawl docs and firecrawl-mcp v3.25.3; keyless endpoint tested live

You ask Claude Code to pull the plans from a pricing page, and it hands back a vague summary instead of the table. Its built-in WebFetch doesn’t run JavaScript, and the agent gets a processed summary rather than the page itself. JS-rendered docs, pricing tables, and “read these 30 pages and compare them” jobs are where it falls apart.

Firecrawl MCP fixes that with one command. It gives the agent a proper web toolkit (~7.5k GitHub stars, maintained by the Firecrawl team). It can search the web, scrape any URL into clean markdown or JSON, map every URL on a site, crawl whole doc sections, and parse PDFs. The part most guides miss is that there’s now a hosted keyless endpoint. Search, scrape, and parse work with no API key, no account, and no Node.js install.

Only need library docs? Our Context7 MCP setup is the better tool for that. Firecrawl is for everything else on the web.

claude mcp add --transport http --scope user firecrawl https://mcp.firecrawl.dev/v2/mcp

You have shell access and the `claude mcp` CLI on this machine. Set up the Firecrawl MCP server (hosted, keyless) and report back. Do not tell me it's done unless step 4 actually confirms it:

1. Run `claude mcp list`. If a server named firecrawl already exists, stop and tell me how it's configured instead of adding a second one.
2. Add the hosted keyless server (no API key, no Node.js needed):
   `claude mcp add --transport http --scope user firecrawl https://mcp.firecrawl.dev/v2/mcp`
3. Run `claude mcp list` again and confirm firecrawl shows as Connected.
4. Tell me to restart Claude Code, then after restart use firecrawl_scrape on https://docs.firecrawl.dev and show me the first 10 lines of the markdown it returns.

Never ask me to paste an API key into this chat. If I want the full tool set later, I'll switch to the sign-in URL myself.
If any step fails, give me the exact error and the command that produced it. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

claude mcp add --transport http --scope user firecrawl https://mcp.firecrawl.dev/v2/mcp

Search the web for the latest Firecrawl release notes and summarize the sources.

claude mcp remove firecrawl

claude mcp add --transport http --scope user firecrawl https://mcp.firecrawl.dev/v2/mcp-oauth

claude mcp add firecrawl --scope user -e FIRECRAWL_API_KEY=fc-YOUR_API_KEY -- npx -y firecrawl-mcp

claude mcp add firecrawl --scope user -e FIRECRAWL_API_KEY=fc-YOUR_API_KEY -- cmd /c npx -y firecrawl-mcp

What Each Piece Does

With sign-in or an API key, the full profile registers 26 tools, including developer search over GitHub issues and PRs, research-paper search, and crawl/agent status checks. The table covers the ones you’ll actually reach for.

Verify It’s Actually Working

claude mcp list

“Connected” only confirms the server answered. The real test is a page WebFetch struggles with, such as a JavaScript-heavy page:

Use firecrawl_scrape on [A JS-HEAVY PAGE, e.g. a pricing page] and give me the exact plan names and prices as a markdown table. Quote them verbatim from the scraped content, don't summarize.

If the tool calls show firecrawl_scrape and the table matches what you see in your browser, it’s working end to end.

Common Mistakes to Avoid

• Opening the MCP URL in a browser. https://mcp.firecrawl.dev/v2/mcp is a server endpoint for your MCP client, not a web page. Sign-in happens through /mcp in Claude Code.
• Pasting your API key into the chat. Firecrawl’s docs are explicit: never put a key in an agent conversation or in the server URL. Use browser sign-in, or pass it with -e / -H in your own terminal.
• Adding a second Firecrawl entry to “upgrade”. Remove the keyless server before adding the sign-in one, or Claude Code ends up with duplicate tools.
• Expecting crawl or map on keyless. Keyless is limited to search, scrape, and parse. If firecrawl_crawl is missing, that’s why, not a broken install.
• Crawling without a limit. Every page costs credits. Run firecrawl_map first, then crawl with limit and includePaths set so a docs crawl doesn’t eat your whole free allowance.

Q&A

Official Resources

• Firecrawl MCP Server on GitHub — source, full tool reference, changelog
• Firecrawl MCP docs — keyless, sign-in, and API-key setup
• Run Firecrawl MCP locally — stdio, Streamable HTTP, self-hosted API
• Rate limits — keyless allowance and plan limits
• Claude Code MCP Documentation — the claude mcp command reference

Last Verified: September 2026 — commands checked against the official ChromeDevTools GitHub docs (v1.10.1)

The page is broken, and you’re the middleman: open DevTools, copy the red console error, paste it into Claude Code, go back for the failed network request, paste again, then explain the layout shift in words. Claude Code can read your source, but it’s debugging blind and guessing from whatever you remembered to copy.

Chrome DevTools MCP takes you out of that loop. It’s the official server from the Chrome DevTools team (~52k GitHub stars), and setup is one command. The agent opens a real Chrome, then reads console messages (with source-mapped stack traces), inspects network requests, records performance traces, runs Lighthouse audits, and takes heap snapshots. It then fixes the code based on what it actually saw.

Already using our Playwright MCP setup? The two are complementary. Playwright is built for driving the browser through flows. DevTools MCP is built for finding out why a page is broken or slow.

claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

You have shell access and the `claude mcp` CLI on this machine. Set up the Chrome DevTools MCP server and report back. Do not tell me it's done unless step 5 actually confirms it:

1. Check `node --version`. It needs a current Node.js LTS (20.19+ or 22.12+). If Node is missing or older, install the current LTS using whatever this machine already uses (nvm, Homebrew, winget, or the official installer from nodejs.org), then confirm `node --version` and `npx --version` work in this shell.
2. Confirm Google Chrome (stable) is installed. If it isn't, stop and tell me. Don't substitute another browser.
3. Run `claude mcp list` to see what's already configured.
4. Add the server:
   - macOS / Linux: `claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest`
   - Windows: `claude mcp add chrome-devtools --scope user -- cmd /c npx -y chrome-devtools-mcp@latest`
5. Run `claude mcp list` again and confirm chrome-devtools shows as Connected.

If any step fails, give me the exact error and the command that produced it. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

node --version

claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

claude mcp add chrome-devtools --scope user -- cmd /c npx -y chrome-devtools-mcp@latest

/plugin marketplace add ChromeDevTools/chrome-devtools-mcp

/plugin install chrome-devtools-mcp@chrome-devtools-plugins

Check the performance of https://developers.chrome.com

claude mcp add chrome-devtools --scope user -- npx chrome-devtools-mcp@latest --isolated

claude mcp add chrome-devtools --scope user -- npx chrome-devtools-mcp@latest --headless

claude mcp add chrome-devtools --scope user -- npx chrome-devtools-mcp@latest --autoConnect

What Each Piece Does

Extensions, PWA, WebMCP, and third-party tool groups also exist but are off by default. Each needs its own flag. See the tool reference below.

Verify It’s Actually Working

claude mcp list

“Connected” only confirms the server started. The real test is a debugging task the agent can’t answer from your source code alone:

Open http://localhost:3000 (or [YOUR APP URL]) in Chrome. List any console errors and any network requests that returned 4xx/5xx, then tell me the most likely cause of each, citing the file and line from the stack trace.

If it comes back with real console output and request statuses (not guesses), it’s working end to end.

Common Mistakes to Avoid

• Thinking it’s broken because Chrome didn’t open. The browser only launches when the agent calls a tool that needs it. Connecting alone never opens a window.
• Skipping the cmd /c wrapper on Windows. Plain npx often fails to launch from inside another process on Windows. That’s the classic “Connection closed” error.
• Installing both the CLI version and the plugin. The docs say to remove the CLI-installed server before installing the plugin, or you’ll have two copies fighting over the same browser profile.
• Browsing sensitive accounts in the agent’s Chrome. The server exposes the browser’s content to your MCP client. With --autoConnect it can see every open window in that profile. Use --isolated when you don’t need your logins.
• Not knowing what gets sent to Google. Usage statistics are on by default, and performance traces can send page URLs to the CrUX API. Add --no-usage-statistics and --no-performance-crux if that matters for your project.

Q&A

Official Resources

• Chrome DevTools MCP on GitHub — source, changelog, issues
• Tool reference — every tool and the flag that enables it
• Configuration guide — all server flags
• Troubleshooting — Windows, WSL, sandbox issues
• Claude Code MCP Documentation — the claude mcp command reference

Last Verified: September 2026 — commands checked against Serena’s official GitHub docs and CLI source

You ask Claude Code to change one function, and it greps for the name, opens a 900-line file, reads half of it, opens three more files, and your context window is gone before the first edit. Then the rename it does is a dozen fragile find-and-replace edits that still miss a call site.

Serena fixes that by giving the agent the tools your IDE already has. It can jump straight to a symbol, list every reference to it, rename it across the project, or replace one function body without reading the rest of the file. Under the hood it runs real language servers (LSP) for 40+ languages. It’s free and open source (by Oraios AI, ~30k GitHub stars, updated daily), and setup is three commands.

uv tool install -p 3.13 serena-agent
serena init
serena setup claude-code

You have shell access and the `claude mcp` CLI on this machine. Set up the Serena MCP server and report back. Do not tell me it's done unless step 5 actually confirms it:

1. Check `uv --version`. If uv is missing, install it with the official installer for my OS:
   - macOS / Linux: `curl -LsSf https://astral.sh/uv/install.sh | sh`
   - Windows (PowerShell): `powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"`
   The installer puts uv in `~/.local/bin`, which may not be on this shell's PATH yet. Add it for this session (or call uv by its full path), then confirm `uv --version` works before continuing.
2. Install Serena: `uv tool install -p 3.13 serena-agent`, then run `serena init`.
3. Register it with Claude Code: `serena setup claude-code`. If that fails, fall back to: `claude mcp add --scope user serena -- serena start-mcp-server --context claude-code --project-from-cwd`
4. From my project root, run `serena project health-check` and show me the output.
5. Run `claude mcp list` and confirm serena shows as Connected.

If any step fails, give me the exact error and the command that produced it. Do not skip ahead.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

curl -LsSf https://astral.sh/uv/install.sh | sh

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

uv tool install -p 3.13 serena-agent

serena init

serena setup claude-code

claude mcp add --scope user serena -- serena start-mcp-server --context claude-code --project-from-cwd

claude mcp add serena -- serena start-mcp-server --context claude-code --project "$(pwd)"

serena project index

claude --system-prompt="$(serena prompts print-cc-system-prompt-override)"

{
  "hooks": {
    "PreToolUse": [
      { "matcher": "", "hooks": [{ "type": "command", "command": "serena-hooks remind --client=claude-code" }] },
      { "matcher": "mcp__serena__*", "hooks": [{ "type": "command", "command": "serena-hooks auto-approve --client=claude-code" }] }
    ],
    "SessionStart": [
      { "matcher": "", "hooks": [{ "type": "command", "command": "serena-hooks activate --client=claude-code" }] }
    ],
    "SessionEnd": [
      { "matcher": "", "hooks": [{ "type": "command", "command": "serena-hooks cleanup --client=claude-code" }] }
    ]
  }
}

What Each Piece Does

The tools the agent gets

Verify It’s Actually Working

claude mcp list

“Connected” only proves the server started. Check that the language server can read your code:

serena project health-check

Then give the agent a task that needs symbol-level understanding, and watch which tools it calls:

Activate the current project with Serena. Then use Serena's tools (not grep) to find every place that calls [A FUNCTION IN YOUR CODE] and summarize how each caller uses it.

If the tool calls show mcp__serena__find_symbol and find_referencing_symbols instead of Grep and Read, it’s working end to end.

Common Mistakes to Avoid

• Installing it from an MCP or plugin marketplace. Serena’s README explicitly warns that marketplace listings carry outdated install commands. Use uv tool install as shown above.
• Assuming “Connected” means “used”. Claude Code often keeps reaching for its built-in grep/read tools. Use the system-prompt override (Step 6) and, for long sessions, the hooks (Step 7).
• Startup timeouts on big projects. If Serena doesn’t come up in time, raise the MCP timeout: export MCP_TIMEOUT=60000 in your shell profile (Windows: setx MCP_TIMEOUT 60000, then open a new terminal).
• serena not found by Claude Code. Even when it works in your terminal, the client may not see it on PATH. Replace serena in the claude mcp add command with the full path (which serena on macOS/Linux, where serena on Windows).
• Forgetting to activate the project with a global install. With --project-from-cwd, start Claude Code from the project root. Otherwise ask it to “activate the current directory as a project using Serena”.

Q&A

Official Resources

• Serena on GitHub — source, README, issue tracker
• Serena client setup docs — Claude Code, Codex, IDEs
• Serena language support — per-language requirements
• uv installation docs
• Claude Code MCP Documentation — the claude mcp command reference

Last Verified: September 2026 — commands and maintenance status checked against live GitHub sources

Claude Code writes the code with confidence, you run it, and it fails: the method was renamed a few versions ago, or the config flag no longer exists. It answers from training data by default, so library APIs, framework syntax, and CLI flags can be months or years out of date.

Context7 fixes that by fetching current, version-matched documentation for whatever library you’re actually using, on demand, instead of relying on what the model memorized during training. Setup takes about 2 minutes.

claude mcp add --scope user --transport http context7 https://mcp.context7.com/mcp

You have access to the `claude mcp` CLI in this project. Do the following and report back — do not tell me it's done unless step 3 actually confirms it:

1. Run `claude mcp list` to see what's already configured.
2. Add Context7: `claude mcp add --scope user --transport http context7 https://mcp.context7.com/mcp`
3. Run `claude mcp list` again and confirm context7 shows as Connected.
4. Prove it works: use the context7 tools to look up current documentation for [NAME A LIBRARY YOU ACTUALLY USE, e.g. "Next.js" or "Prisma"] and summarize one recent API detail. If step 3 or 4 fails, tell me the exact error instead of reporting success.

What You’ll Need

Step-by-Step Setup

Prefer to do it by hand instead of delegating to the agent? Here’s the manual version.

claude mcp add --scope user --transport http context7 https://mcp.context7.com/mcp

claude mcp add --scope user --header "CONTEXT7_API_KEY: YOUR_API_KEY" --transport http context7 https://mcp.context7.com/mcp

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

What Each Piece Does

Verify It’s Actually Pulling Fresh Docs

claude mcp list

“Connected” only confirms the handshake. The real test is asking about a library you know has changed recently and checking the answer reflects that:

What's the current recommended way to set up middleware in [a framework you use]? Use Context7 to check current docs before answering, and tell me which version the docs you pulled are for.

If the answer cites a specific version and matches what’s actually in that library’s current docs (not what an older training cutoff would guess), it’s working end to end.

Common Mistakes to Avoid

• Assuming it’s used automatically every time. Claude decides when a query needs current docs versus when its own knowledge is sufficient — if you want to force it, say “use Context7” explicitly in the prompt.
• Hitting rate limits on the free unauthenticated tier. Add an API key (Step 2) if you’re using it heavily — it’s free and takes about a minute.
• Confusing this with a general web search tool. It’s scoped to library/framework/API documentation, not general current events or arbitrary web content.
• Not restarting after adding the server. Claude Code needs a fresh session or an /mcp panel refresh to pick up a newly added server.

Q&A

Official Resources

• Context7 Documentation — setup guides for every MCP client
• Context7 Dashboard — create a free API key
• Claude Code MCP Documentation — the claude mcp command reference