The MCP server ecosystem exploded. Thousands of servers now exist in the registry, and the number grows every week. Most of them are toys. Some of them are transformative. The difference between a productive Claude Code setup and a cluttered one is knowing which servers actually earn their place in your configuration.
We run Claude Code as our primary development environment at Like One. Not as an experiment — as the production tool that ships our software, manages our infrastructure, and generates our content. We have tested dozens of MCP servers across hundreds of sessions. This is the list of servers that survived contact with real work.
If you are new to MCP, start with our explainer on what Model Context Protocol is and why it matters. If you already know the protocol and want to build your own, jump to our MCP server tutorial. This guide assumes you have Claude Code installed and want to know which servers to add to your .mcp.json.
How MCP Servers Work in Claude Code
MCP servers are external processes that expose tools, resources, and prompts to Claude through a standardized protocol. When you add a server to your Claude Code configuration, Claude discovers its capabilities at startup and can use them during any session. No wrapper code. No custom integrations. You declare the server, Claude handles the rest.
Configuration lives in .mcp.json at your project root or ~/.claude/mcp.json for global servers. Each entry specifies how to launch the server:
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {
"GITHUB_PERSONAL_ACCESS_TOKEN": "your-token"
}
}
}
}Claude Code launches each server as a subprocess, communicates over stdio, and manages the lifecycle automatically. You can run multiple servers simultaneously — the tools from all active servers appear in Claude's tool palette. The practical limit is around 10 to 15 servers before tool discovery latency becomes noticeable.
How We Evaluate MCP Servers
Not all MCP servers deserve your configuration file. Before adding any server, we evaluate it on five criteria:
- Does it solve a real problem? If you can accomplish the same thing with a bash command in less time, the server adds complexity without value.
- Is it maintained? Servers with no commits in 90 days are abandoned software. MCP is evolving fast — unmaintained servers break.
- Is it secure? Every MCP server gets access to your Claude session. A malicious or poorly written server can exfiltrate data, execute arbitrary code, or corrupt your workflow. Audit before you trust.
- Does it respect scope? A good server does one thing well. A server that tries to be a filesystem manager, a database client, and a web scraper is three bugs waiting to happen.
- Does it actually work with Claude Code? Some servers are built for other MCP clients and break in Claude Code's execution model. We only list servers we have run successfully.
The 15 Best MCP Servers for Claude Code
1. Filesystem Server
Package: @modelcontextprotocol/server-filesystem
The official filesystem server from Anthropic gives Claude scoped access to directories you specify. Read files, write files, search directories, and move files — all within the boundaries you define. Unlike Claude Code's built-in file tools, the filesystem MCP server lets you expose specific directories to Agent SDK agents without giving them access to your entire system.
Best for: agent architectures where you need controlled filesystem access, multi-project setups where Claude should only see specific directories, and sandboxed execution environments.
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/project"]
}2. GitHub Server
Package: @modelcontextprotocol/server-github
The official GitHub MCP server gives Claude full access to repositories, pull requests, issues, branches, and code search through the GitHub API. Create branches, open PRs, review diffs, comment on issues, and manage releases without leaving your Claude Code session.
This is the single most useful MCP server for development workflows. Instead of switching to the browser to check a PR or read an issue, Claude reads it directly and can act on the information immediately. Combined with agentic loops, you get automated code review, issue triage, and release management.
Best for: any developer using GitHub. Period.
3. PostgreSQL Server
Package: @modelcontextprotocol/server-postgres
Connects Claude to a PostgreSQL database for schema inspection, query execution, and data analysis. Claude can explore your database structure, write and run SQL queries, and analyze results — all within the conversation context.
The server enforces read-only access by default, which is the right default for a tool that an AI model controls. You can enable write access, but think twice. A production database connected to an autonomous agent with write permissions is a decision you should make deliberately, not accidentally.
Best for: data analysis, database debugging, schema exploration, and generating migration scripts based on current state.
4. SQLite Server
Package: @modelcontextprotocol/server-sqlite
The lightweight alternative to the PostgreSQL server. Connects to local SQLite databases for querying, schema inspection, and data manipulation. Particularly useful for development databases, application caches, and embedded data stores.
We use this extensively with our sovereign brain architecture, which stores over 900 knowledge entries in SQLite with vector search. Claude can query the brain database directly to understand stored knowledge, debug retrieval issues, and verify data integrity.
Best for: local development, embedded databases, prototyping, and any project using SQLite as its data layer.
5. Playwright Browser Server
Package: @anthropic/mcp-server-playwright
Gives Claude control of a real browser through Playwright. Navigate to pages, take screenshots, fill forms, click buttons, extract content, and run end-to-end tests. This is not a simple HTTP fetcher — it runs a full Chromium instance that handles JavaScript rendering, authentication flows, and dynamic content.
The difference between this and a basic web fetch tool is the difference between reading a recipe and cooking the meal. Claude can interact with web applications the way a human does, which unlocks automation workflows that API-only approaches cannot touch.
Best for: web scraping, end-to-end testing, form automation, visual regression testing, and any workflow that requires interacting with a web UI.
6. Slack Server
Package: @anthropic/mcp-server-slack
Connects Claude to Slack for reading channels, sending messages, searching conversations, and managing threads. Claude can monitor channels for questions, draft responses, search message history for context, and post updates — all without you touching the Slack interface.
The productivity gain is real: instead of context-switching between your IDE and Slack every time someone asks a question, Claude reads the question, finds the answer in your codebase, and drafts a response. You review and send. Or if you trust the agent enough, it sends directly.
Best for: team communication, automated status updates, channel monitoring, and reducing context switches.
7. Notion Server
Package: @anthropic/mcp-server-notion
Provides read and write access to Notion workspaces. Claude can search pages, read content, create new pages, update existing ones, and query databases. If your team uses Notion for documentation, project management, or knowledge bases, this server makes Claude a native participant in your Notion workflow.
Best for: documentation management, project planning, knowledge base maintenance, and syncing code changes with project documentation.
8. Memory Server
Package: @modelcontextprotocol/server-memory
A persistent key-value memory store that survives across Claude Code sessions. Claude can save facts, decisions, preferences, and project context, then retrieve them in future sessions. This is the simplest path to giving Claude persistent memory without building custom infrastructure.
The server stores data as a local JSON knowledge graph with entity-relationship modeling. It is not as sophisticated as a full vector search system, but it solves the most painful problem — Claude forgetting everything between sessions — with zero configuration.
Best for: long-running projects where context persistence matters, personal preferences, project-specific knowledge, and any workflow where you are tired of re-explaining things.
9. Fetch / Web Search Server
Package: @modelcontextprotocol/server-fetch
A clean HTTP fetch server that lets Claude retrieve web content, read documentation, and access APIs. It handles HTML-to-markdown conversion, follows redirects, and respects robots.txt. Unlike the browser server, this is lightweight and fast — no browser instance, no JavaScript rendering, just HTTP requests and content extraction.
Best for: reading documentation, fetching API responses, downloading data files, and any task where you need web content without browser interaction.
10. Supabase Server
Package: Available through Claude Code's built-in MCP integration
If you are building on Supabase, this server gives Claude direct access to your project — database queries, schema management, edge function deployment, migration management, and log inspection. Claude can debug production issues by reading logs, inspect your database schema, and even deploy edge functions without leaving the session.
The tight integration between Supabase's MCP server and Claude Code makes it one of the most polished MCP experiences available. Schema changes, migration generation, and type generation all work seamlessly.
Best for: any project using Supabase for backend infrastructure.
11. Sentry Server
Package: @sentry/mcp-server-sentry
Connects Claude to your Sentry error tracking. Claude can search for errors, read stack traces, analyze error frequency, and suggest fixes based on the actual runtime errors your application is producing. Instead of you copying a stack trace from Sentry into Claude, the agent reads it directly and cross-references with your codebase.
Best for: debugging production errors, error triage, and connecting runtime failures to code changes.
12. Vercel Server
Package: Available through Claude Code's built-in MCP integration
Full access to your Vercel deployments, build logs, runtime logs, and project configuration. Claude can check deployment status, read build errors, inspect runtime logs, and even deploy new versions. If you deploy to Vercel, this eliminates the dashboard tab you keep open.
Best for: deployment management, build debugging, and monitoring frontend applications on Vercel.
13. MCP Shield (Security Auditing)
Package: mcp-shield — open source
Built by our team at Like One. MCP Shield scans code, files, and directories for security vulnerabilities using 20 detection rules. It catches hardcoded secrets, SQL injection patterns, command injection risks, insecure configurations, and other OWASP Top 10 issues directly inside Claude Code.
The value is integration: instead of running a separate security scanner and then bringing results back to Claude, the agent scans as part of its normal workflow. Write code, scan it, fix issues — all in one loop. This is how MCP server security should work.
Best for: security-conscious development, pre-commit scanning, code review automation, and compliance workflows.
14. Claude Brain (Persistent Knowledge)
Package: claude-brain — open source
Also built by our team. Claude Brain provides four MCP tools for persistent knowledge management: boot (load all context at session start), search (semantic keyword matching), write (store new knowledge), and read (retrieve specific keys). It turns Claude Code into a system that accumulates knowledge across sessions rather than starting fresh every time.
This is the MCP server that powers the brain architecture described in our persistent memory guide. If you want Claude to remember your project architecture, your coding preferences, your deployment procedures, and your team conventions — this is how.
Best for: long-term projects, team knowledge management, and building agents that learn from experience.
15. lo-eyes (Visual Frontend Inspection)
Package: lo-eyes — open source
Another Like One project. lo-eyes gives Claude the ability to take screenshots of web pages, scan for visual issues, run accessibility audits, and test across device viewports. Four tools: screenshot capture, visual scanning, accessibility auditing, and device simulation.
Frontend development without visual feedback is guesswork. lo-eyes closes the loop — Claude writes CSS, takes a screenshot to verify the result, identifies layout issues, and fixes them in the same session. For accessibility testing, it catches contrast failures, missing alt text, and touch target violations automatically.
Best for: frontend development, accessibility compliance, visual QA, responsive design testing, and any workflow where Claude needs to see what it built.
Servers We Evaluated and Skipped
Not every popular server made the list. Some notable exclusions:
- Generic AI wrapper servers that just proxy to other LLMs. Claude does not need another LLM in the loop for most tasks. Added latency, added cost, marginal value.
- Kitchen-sink servers that expose 50+ tools. Claude performs worse with too many tools — it spends more tokens on tool selection than execution. Prefer focused servers.
- Servers without authentication controls. If a server connects to an external service and does not let you scope its permissions, it is a security liability. We skip these regardless of functionality.
- Unmaintained servers. The MCP spec is still evolving. Servers that have not been updated in months likely have compatibility issues that will waste your time.
Configuration Best Practices
After running MCP servers across hundreds of sessions, these patterns consistently produce better results:
Use project-level configuration for project-specific servers. Put database connections, project APIs, and team tools in .mcp.json at the project root. Put personal tools like memory, browser, and general-purpose servers in ~/.claude/mcp.json.
Set environment variables in the config, not in your shell. MCP server configuration supports env blocks. Use them. This keeps credentials scoped to the server that needs them and prevents token leakage to other processes.
Start with three servers, not thirteen. Add servers one at a time as you encounter real needs. GitHub, filesystem, and memory cover most development workflows. Add specialized servers when you hit a specific wall, not because the server exists.
Audit third-party servers before installing. Read the source code. Check what permissions it requests. Verify it does not phone home or collect telemetry. A compromised MCP server has the same access as Claude itself — treat it with the same caution you would treat a VS Code extension with full filesystem access.
Monitor server health. MCP servers crash. They hang. They lose connection. Claude Code handles reconnection gracefully in most cases, but if a server is consistently unstable, remove it. A broken server that Claude keeps trying to use wastes more time than having no server at all.
Building Your Own MCP Server
The best MCP server is often the one you build for your specific workflow. If you have an internal API, a custom database, or a proprietary tool that your team uses daily, wrapping it in an MCP server makes it natively accessible to Claude.
Building an MCP server takes less time than you think. The protocol handles the hard parts — serialization, tool discovery, connection management. You write the business logic. A basic server with three tools can be built in under an hour.
We have a complete walkthrough in our MCP server tutorial that takes you from zero to a working server. If you want to understand the protocol itself before building, read our MCP explainer first.
The MCP Ecosystem Is Still Early
Thousands of MCP servers exist today. Most are experiments. Some are production-grade. The 15 listed here are the ones that have proven their value across real development workflows, not demo sessions.
The protocol is stabilizing, but it is not frozen. Expect breaking changes, new capabilities, and shifts in which servers matter as the ecosystem matures. The evaluation criteria matter more than the specific list — apply them to every new server you consider, and you will avoid the trap of installing everything and getting value from nothing.
MCP servers are not plugins. They are capabilities. Each one you add makes Claude more useful at a specific task. Choose the ones that match your actual work, ignore the rest, and build your own when the ecosystem does not serve your needs.
For security guidance on evaluating servers, read our MCP security audit checklist. For connecting servers to autonomous agents, see our Claude Agent SDK guide. And for the fundamentals of how agents use these tools, start with our explanation of agentic loops.