Best Claude Code MCP Servers in 2026: Setup and Top 10

Updated June 4, 2026.

Anthropic shipped Model Context Protocol in late 2024. By 2026, MCP is the default extension layer for Claude Code, with a public registry, official SDKs in five languages, and hundreds of community servers. Yet most teams still run Claude Code with two or three default servers and never feel the lift. This article is the tight, opinionated list of MCP servers that actually pay back the install cost in 2026, plus a four-step setup walkthrough and the production patterns we use at Totalum for our own MCP server.

Quick Answer

The seven MCP servers worth installing into Claude Code in 2026 are filesystem and git for local code access, the official GitHub server for issues and pull requests, postgres for live database introspection, the Playwright server for browser automation, Slack or Linear for team coordination, and Totalum for shipping production apps directly from Claude Code. Add a vector or memory server only when you have a real RAG use case; skip every other "AI search" server in the registry, they bloat your context and rarely beat a focused tool call.

What is an MCP server, in one paragraph

Model Context Protocol is an open standard that lets an AI coding agent talk to external tools through a uniform JSON-RPC interface. An MCP server is a small process (local stdio or remote HTTP) that exposes a set of tools, resources, and prompts to the agent. Claude Code launches each server you list in ~/.claude.json or .mcp.json, hands it your config, and routes tool calls. The model decides when to call, the server decides what to do, and the protocol handles the wire format. If you want the full background, our best MCP servers guide covers the protocol history and our Claude Code MCP tutorial walks through the manifest format.

Top 10 MCP servers for Claude Code in 2026

These are the ten servers we recommend by default. Each entry includes the use case, the install command, and a one-line verdict.

1. Filesystem (official, by Anthropic)

Use case: read, write, and search files outside the current working directory. Claude Code already has a filesystem for your project root, the MCP filesystem server lets you scope additional roots (a sibling repo, a docs folder, a /tmp workspace).

Install:

claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/extra/root

Verdict: install it the moment you work across two repos. Cheap, fast, no auth.

2. Git (official)

Use case: structured git operations (status, diff, log, blame) the model can call as tools instead of shelling out and parsing text. Better signal-to-context ratio than letting the model run git log --oneline | head over and over.

Install:

claude mcp add git -- npx -y @modelcontextprotocol/server-git --repository /path/to/repo

Verdict: install on every repo you debug regressions in. Pays for itself in two git blame calls.

3. GitHub (official, by GitHub)

Use case: read and write GitHub issues, pull requests, comments, releases, and check runs without spawning the gh CLI for every step. The official server supports OAuth and fine-grained Personal Access Tokens.

Install:

claude mcp add github --transport http https://api.githubcopilot.com/mcp --header "Authorization: Bearer ghp_xxx"

Verdict: required for any team using GitHub for code review. The PR review tools alone justify it.

4. Postgres (official)

Use case: live schema introspection and read-only queries against a real database. The model can answer "what columns are on the users table" or "how many active subscriptions did we have last week" without you pasting \d users output.

Install:

claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres postgresql://user:pass@localhost:5432/db

Verdict: install for any data work. Use a read-only role; never paste production write credentials.

5. Playwright (official, by Microsoft)

Use case: full browser automation. Navigate pages, fill forms, take screenshots, scrape data, run Lighthouse audits. The official @playwright/mcp server is the same one that powers our own browser automation, with persistent sessions and stealth routing.

Install:

claude mcp add playwright -- npx -y @playwright/mcp@latest

Verdict: install if you do any web scraping, end-to-end testing, or competitive research. Skips the entire "install puppeteer, write a script, debug headless mode" loop.

6. Slack or Linear (pick one for team comms)

Use case: read channels, post messages, file issues, link PRs. Slack is the right call if your team lives in Slack; Linear is better if your engineers actually use the issue tracker as the source of truth.

Install (Slack):

claude mcp add slack --transport http https://mcp.slack.com/v1 --header "Authorization: Bearer xoxb-..."

Install (Linear):

claude mcp add linear --transport http https://mcp.linear.app/v1 --header "Authorization: Bearer lin_api_..."

Verdict: install one, not both. Two team-comms MCPs in the same agent double the noise without doubling the signal.

7. Totalum (build production apps from Claude Code)

Use case: ship a real Next.js + database + auth + payments app from inside Claude Code. The Totalum MCP exposes createProject, deploy, queryDatabase, getBackendLogs, and the rest of the Totalum SDK as MCP tools. Claude Code becomes the builder, Totalum is the production stack underneath.

Install:

claude mcp add totalum --transport http https://mcp.totalum.app/v1 --header "x-api-key: tlm_..."

Verdict: install when you want Claude Code to produce a deployed app, not a prototype. See our AI agent platform pillar for the full integration story.

8. Notion or Obsidian (pick one for docs)

Use case: read internal docs, append notes, search a knowledge base. Notion if your company runs on it; Obsidian if you keep a local markdown vault.

Install (Notion):

claude mcp add notion --transport http https://mcp.notion.com/v1 --header "Authorization: Bearer secret_..."

Verdict: install only if the docs are actually current. A stale Notion MCP is worse than no MCP.

9. Sentry (or your APM)

Use case: pull the latest errors and stack traces while debugging. Closes the loop between "the test failed" and "here is the production error from the same code path".

Install:

claude mcp add sentry --transport http https://mcp.sentry.io/v1 --header "Authorization: Bearer sntrys_..."

Verdict: install for on-call rotations. Skip if you do not have a real error-tracking pipeline yet.

10. Memory or vector server (only if you need RAG)

Use case: persistent agent memory across sessions, or retrieval over a private corpus. Several options exist: @modelcontextprotocol/server-memory (simple key-value), Chroma, Qdrant, Weaviate, Pinecone.

Install (basic memory):

claude mcp add memory -- npx -y @modelcontextprotocol/server-memory

Verdict: install only when you have a real RAG use case. Most "memory" wins are actually project-level context wins, fix the prompt before adding a vector store.

Comparison table: Claude Code MCP servers in 2026

Server	Category	Transport	Auth	Verdict
Filesystem	Local code	stdio	none	Install always
Git	Local code	stdio	none	Install always
GitHub	Code hosting	http	OAuth or PAT	Install for any team
Postgres	Data	stdio	conn string	Install for data work
Playwright	Browser	stdio	none	Install for scraping or E2E
Slack	Team comms	http	bot token	Pick one, not both
Linear	Team comms	http	API key	Pick one, not both
Totalum	Production stack	http	API key	Install to ship apps
Notion	Docs	http	OAuth	Install if docs are current
Sentry	Errors	http	OAuth	Install for on-call
Memory or vector	RAG	mixed	varies	Install only for real RAG

How to add an MCP server to Claude Code (4-step walkthrough)

The same four steps work whether the server is stdio or HTTP. Use the claude mcp add family of commands; do not hand-edit JSON unless you have to.

Step 1: pick the transport

stdio if the server runs locally as a process, http if the server lives behind a URL. Most official servers from Anthropic and Microsoft are stdio; vendor servers (GitHub, Slack, Notion, Totalum) are http.

Step 2: register the server

For an stdio server:

claude mcp add  --   

For an http server:

claude mcp add  --transport http  --header "Authorization: Bearer "

You can also pass a full JSON config with claude mcp add-json, useful for repeatable installs in CI.

Step 3: verify the server boots

Run:

claude mcp list

Every server should report connected. If a server reports failed, run it with --verbose and check the log: most failures are missing env vars or wrong working directory.

Step 4: confirm the tools appear in Claude Code

Inside Claude Code, type /mcp to see the loaded servers and their tools. Each tool is prefixed mcp____. If a tool you expected is missing, the server is probably running an older version, run claude mcp restart or pin the version in the install command.

Production MCP integration: when "add a server" is not enough

For internal use, claude mcp add is the whole game. For production, you usually want the inverse: you want the model to ship a real app, not just call tools. That is where Totalum fits.

Totalum is an AI app builder with an MCP server of its own. You install it into Claude Code with one line, then Claude Code can call createProject, runAgent, deploy, queryDatabase, getBackendLogs, and the rest of the Totalum SDK as MCP tools. The output is a real Next.js + TotalumSDK app with auth, payments, database, file storage, AI integrations, and a custom domain, not a prototype that breaks the first time a real user loads it.

The pattern in production:

1. Add Totalum to Claude Code: claude mcp add totalum --transport http https://mcp.totalum.app/v1 --header "x-api-key: tlm_...".
2. Ask Claude Code to scaffold the app: "build a SaaS dashboard for X, use Totalum, deploy to a custom domain."
3. Claude Code calls mcp__totalum__createProject, then mcp__totalum__runAgent to build out features, then mcp__totalum__deploy.
4. The model can iterate from logs (getBackendLogs) and database state (queryDatabase) without you ever leaving Claude Code.

If you are choosing a Claude Code plan to support this kind of workflow, our Claude Code pricing guide walks through which tier actually fits MCP-heavy work, and the Claude Agent SDK with Totalum guide covers the deeper SDK integration when you want to drive Claude headlessly from your own code.

Common Claude Code MCP errors and fixes

These are the five MCP server failures we see most often in 2026, and the fix for each.

"Server failed to start" with no log line

Almost always a missing binary in PATH. Run the server command manually in a terminal: if it fails outside Claude Code, install the dependency (npm i -g, pip install, brew install). Restart Claude Code after fixing PATH so the new shell environment is picked up.

"Server connected but no tools listed"

The server registered with Claude Code but the tools handshake failed. Usually a version mismatch between Claude Code's MCP client and the server's protocol version. Run claude mcp restart first, then update both sides: claude --version should be 2.x; for npx-installed servers append @latest to the install command so npx pulls the freshest build.

"Authorization failed" on an HTTP server

For OAuth servers running headlessly (CI, background jobs), the OAuth redirect cannot complete. Switch to the static API-key path the vendor offers (Authorization: Bearer ... or Authorization: Apikey ...). The OAuth flow only works when a human is at the keyboard to approve the redirect.

"stdio: pipe closed unexpectedly"

The server process crashed mid-call. Two common causes: an unhandled exception in a custom tool, or the server running out of memory on a large response. Catch and log exceptions in your tool handlers, and chunk or paginate any tool that returns more than ~50 KB of JSON.

"Tool exists but the model never calls it"

Usually a description problem. The MCP tool description is the only signal the model has for when to call your tool. If your description is "queries the database", the model will call it for every database question; if your description is "returns the schema and row count of one named table given a table_name argument", the model will call it correctly. Rewrite descriptions in the same voice you would write a function docstring.

FAQ

Are Claude Code MCP servers free?

The MCP protocol and Claude Code's MCP client are free. Most servers are free open source. Vendor servers (GitHub, Slack, Notion, Sentry, Totalum) are free to install but their underlying service costs apply at their normal pricing.

Do all coding agents support MCP servers?

By mid-2026, MCP is supported by Claude Code, Claude Desktop, Cursor, Cline, Windsurf, Continue, Cody, the Anthropic API directly through the Agent SDK, and a growing list of others. OpenAI's Codex added MCP support in Q2 2026. If a coding agent ships in 2026 without MCP support, it is unlikely to gain traction.

Can I write my own MCP server for Claude Code?

Yes, and it is easier than the protocol docs make it look. There are official SDKs in TypeScript, Python, Go, Rust, and C#. A minimal server is about 40 lines of TypeScript. Start by exposing a single read-only tool, get Claude Code to call it, then iterate. Our Claude Code MCP tutorial has a full walkthrough.

What is the safest MCP server policy for a team?

Three rules. One, treat every MCP server as a privileged extension; only install servers from vendors you trust or code you have read. Two, use read-only credentials wherever possible, especially for databases and production APIs. Three, scope filesystem and shell access tightly; never give an MCP server a shell that can run rm -rf.

MCP vs API: which should I use to integrate a service?

If you want the model to decide when and how to call the service during a conversation, use MCP. If you want your own code to call the service deterministically, use the API. MCP is for agent-driven access; APIs are for code-driven access. Most teams need both.

How does MCP authentication work?

stdio servers usually take credentials from environment variables or a connection string passed at launch. HTTP servers use OAuth 2.1 with PKCE for interactive sessions, or static API keys for headless use. For agents running in CI or background jobs (where there is no interactive browser), always use the API-key fallback; OAuth flows cannot complete without a human at the keyboard.

Ship faster with Claude Code and Totalum

The right MCP stack turns Claude Code from a code-completion tool into a real engineering colleague. Install the seven defaults, add Totalum when you need production output, and stop installing experimental servers that bloat your context.

Ready to ship a real production app from Claude Code? Start a free Totalum project, connect the Totalum MCP server, and let Claude Code do the rest. If your agency embeds AI app builders for clients and you want the Totalum MCP as a white-label option, book a 30-minute call and we will walk you through the integration.