MCP Server Roundup: Ten You Can Actually Install

The list first: ten entries, install commands below

How to install (commands copied from each project's own page)

# 4 Browser: this exact line is in the official README
claude mcp add playwright npx @playwright/mcp@latest

# 3 Chrome DevTools: the project ships a JSON config instead
{"mcpServers":{"chrome-devtools":{"command":"npx","args":["-y","chrome-devtools-mcp@latest"]}}}

# 2 Context7: the README now points at this installer (--claude targets the client)
npx ctx7 setup

# 7 Exa: the repo points at the Claude plugin marketplace
claude plugin install exa@claude-plugins-official

# 8 Notion / 9 Supabase: remote HTTP, OAuth, no secrets in your config file
claude mcp add --transport http notion https://mcp.notion.com/mcp
claude mcp add --transport http supabase https://mcp.supabase.com/mcp

# 1 Official reference servers (two examples)
npx -y @modelcontextprotocol/server-filesystem /your/allowed/dir
uvx mcp-server-git

# Do these two things the moment you finish installing
/context   # see what it costs
/mcp       # switch off what today's work does not need

Entry by entry, in plain words

• 1. modelcontextprotocol/servers — seven reference servers remain: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time. Filesystem's whole value is its permission boundary: it refuses to start without at least one allowed directory, and clients supporting roots can replace that list dynamically. But on Claude Code, Filesystem and Git are largely redundant — Read/Edit and plain git already cover them. The repo's real use is as a textbook: this is how the maintainers themselves draw a tool boundary.
• 2. upstash/context7 — fetches the current official docs for a library on demand, treating the model's training cutoff as the chronic illness it is. Second-highest star count here, and the most over-installed. Skill 10 digs into it properly, including when pasting the doc yourself simply beats it.
• 3. ChromeDevTools/chrome-devtools-mcp — the project's own line is that it "lets your coding agent control and inspect a live Chrome browser." The split from entry 4 is clean: this one records performance traces, inspects network requests, and reads console messages with source-mapped stack traces. Playwright MCP does none of those. Debugging "why is this page slow" is its territory.
• 4. microsoft/playwright-mcp — official description: it lets models "interact with web pages through structured accessibility snapshots, bypassing the need for screenshots or visually-tuned models." Structured snapshots are both cheaper in tokens and steadier than pixels — the subject of Skill 11. Install it to drive flows, fill forms, and assert.
• 5. github/github-mcp-server — a hosted remote instance lives at https://api.githubcopilot.com/mcp/. What matters is not the feature list but how you narrow it: the docs list 100+ tools grouped into context / repos / issues / pull_requests / users toolsets, selectable with --toolsets, and --read-only drops every write tool at once. Installing it unnarrowed is the most expensive click on this list.
• 6. modelcontextprotocol/registry — the official community registry, described as an app store for MCP servers, with its API at registry.modelcontextprotocol.io. It is still in preview, its v0.1 API is frozen, and data resets remain possible — which makes it a place to check whether a name is real, not a shopping list. On the Claude side there is also the human-reviewed Anthropic Directory.
• 7. exa-labs/exa-mcp-server — described as connecting "AI agents to Exa for web search, content fetching, and multi-step research." The one good reason to install it is that your harness has no built-in web search. If it has one, skip: two search tools online at once fight over every trigger.
• 8. makenotion/notion-mcp-server — the official Notion server, and it carries a detail worth stealing: it ships retrieve-page-markdown and update-page-markdown, reading and writing pages as Markdown rather than block JSON, which the project states is significantly more token-efficient for AI agents. That is what a well-built server looks like — the author thought about the size of what comes back.
• 9. supabase-community/supabase-mcp — hosted at https://mcp.supabase.com/mcp, with two switches worth knowing: readOnly excludes every mutating tool, and projectScoped omits project_id from tool schemas and drops account-level tools. Both belong on whenever a production project is in reach.
• 10. getsentry/sentry-mcp — hosted at https://mcp.sentry.dev, installable on Claude Code with claude plugin install sentry-mcp@sentry-mcp. It moves live errors, stack traces and blast radius into the session where you are already fixing the bug. It has 845 stars — the lowest count here, and also the most narrowly scoped first-party server here. That is the argument against ranking by stars, in one row.

Why these ten cannot share a ranking: four lines and one trap

The four lines

1. Docs retrieval (2, 7) — cures "the model does not know the new version." Input is a question, output is text: read-only, no side effects, and a failure costs a few hundred wasted tokens at worst. Lowest-risk line, and the easiest to install twice by accident.
2. Browser (3, 4) — cures "the model cannot see what rendered." It has side effects (it really clicks, really types), its returns can be enormous, and the web pages it reads back are untrusted input. Entry 3 leans diagnostic (performance, network, console); entry 4 leans operational (accessibility snapshots, click, fill, assert). Decide which you are before installing.
3. Code hosting (5) — cures "the model cannot see the PR or CI." Most tools, most sensitive permissions, and the only line where "how do I narrow this" comes before "should I install this." Note that the official docs themselves recommend the opposite move: a CLI like gh is more context-efficient than an MCP server precisely because it adds no per-tool listing.
4. Business systems and office suites (8, 9, 10) — cures "the model cannot reach the systems your company runs on." What they share is remote HTTP plus OAuth: no credentials in your config file, but your data leaves your machine. Pick on whether return size is controllable and whether a read-only mode exists, not on the feature matrix.

And one thing that belongs to no line

The official reference servers inside entry 1 — Filesystem, Git, Memory, Sequential Thinking, Time — are protocol demos, not products. They exist so you can see what a server ought to look like. On a coding harness that already has file tools and a shell, installing them hands the model a duplicate tool listing and bills you for it on every turn.

Three-minute inspection: six tests, with these ten as the yardstick

1. Is it first-party, and is the repo archived? Read the owner: github/, microsoft/, ChromeDevTools/, makenotion/ and getsentry/ are first-party; nothing under modelcontextprotocol/servers-archived gets installed. Star count plays no part in this test.
2. Is there a read-only switch? --read-only on github-mcp-server and readOnly on supabase-mcp are the positive samples. A server that can mutate your data and offers no read-only mode has a design defect, not a rich feature set.
3. Can tools be selected by group? Granularity like --toolsets context,repos decides your tax bill. A server that mounts 100+ tools at once and gives you no way to trim is building its convenience on your context window.
4. Does the author care about return size? Notion's two *-markdown tools are the benchmark. The opposite is returning a whole block JSON tree, a whole table, a whole trace — on Claude Code, a single MCP tool result is capped at 25,000 tokens by default and triggers a warning above 10,000, with oversized output written to a file the model then has to open.
5. Local stdio or remote HTTP, and where do the credentials live? Local: one more resident process and its startup latency, but data stays on your machine. Remote: OAuth spares you key management, but data leaves. Among the three claude mcp add scopes, --scope project writes into the repo's .mcp.json and shares it with the whole team — never put personal credentials there.
6. Is there a CLI that already does this? The docs say it outright: gh, aws, gcloud and sentry-cli are more context-efficient than MCP servers because they add no per-tool listing at all — the model just runs the command. If a CLI solves it, do not connect a server.

The honest cost: where the context tax is actually collected

The good news first

The Claude Code docs are explicit: MCP tool definitions are deferred by default, so only tool names and server instructions enter context until Claude actually uses a specific tool. The old rule of thumb that ten servers burn half your window has expired. To see what they really cost right now, run /context; to switch one off for the session, run /mcp.

Now the part that did not change

• Return values are the real bill. Anthropic's "Code execution with MCP" post benchmarked a Google Drive to Salesforce workflow: rewriting it so the agent wrote code against its tools instead of calling them one at a time dropped token usage from 150,000 to 2,000, a 98.7% reduction (per the Anthropic engineering blog, 2025-11). Most of those 148,000 tokens were intermediate results passing through context verbatim. That is why test 4 carries the weight in this issue.
• Server instructions are not deferred. Definitions load lazily; the server's own instruction block stays resident. Ten servers means ten resident instruction blocks.
• Trigger conflicts are unchanged. Two search tools or two browser tools online at once force a choice on every turn — and choosing wrong costs far more than a few hundred tokens. That is the real reason behind "skip entry 7 if your harness already searches."
• Untrusted return values are an injection surface. The docs put it plainly: verify you trust each server before connecting it, because servers that fetch external content expose you to prompt injection risk. The browser and docs-retrieval lines deserve the most suspicion — their entire job is carrying strangers' text into your context. This is the subject of Skill 12.
• Installed-but-unused is pure loss. Claude Code's /usage breaks recent usage down by skill, subagent, plugin and individual MCP server. Check it fortnightly; anything with a visible share and no memory of what it did for you comes off.