The 21 Best Claude Code MCPs for Developers in 2026
The 21 most powerful MCP servers for Claude Code — connect databases, APIs, and project management tools to make Claude Code far more capable.
How to Install MCPs
Complete MCP Reference Table
Top 5 MCPs: Detailed Use Cases and Configuration
A Real .mcp.json Configuration for a Production Project
Category 1: Database & Data Access
Category 2: Project Management & Collaboration
Category 3: Development & DevOps
Category 4: Documentation & Knowledge
Category 5: Testing & Quality
Category 6: AI & Specialized Tools
Building Your MCP Stack
Common MCP Mistakes (and How to Avoid Them)
MCP Performance Tips
The Future of MCPs
Deep Dive: PostgreSQL MCP
Deep Dive: GitHub MCP
Deep Dive: Linear MCP
Deep Dive: Sentry MCP
Deep Dive: Context7 MCP
1. PostgreSQL MCP
2. Supabase MCP
3. Redis MCP
4. Linear MCP
5. GitHub MCP
6. Slack MCP
7. Notion MCP
8. Docker MCP
9. Kubernetes MCP
10. AWS MCP
11. Vercel MCP
12. Web Search MCP
13. Confluence MCP
14. Context7 MCP
15. Playwright MCP
16. Sentry MCP
17. Lighthouse MCP
18. Figma MCP
19. OpenAPI MCP
20. Memory MCP
21. Filesystem MCP (Enhanced)
Solo Developer Stack (5 MCPs)
Startup Team Stack (8 MCPs)
Enterprise Stack (12+ MCPs)
Mistake 1: Loading Too Many MCPs at Once
Mistake 2: Hardcoding Credentials in Configuration
Mistake 3: Using Production Credentials for Database MCPs
Mistake 4: Not Testing MCP Connectivity Before Relying on It
Mistake 5: Ignoring MCP Server Logs
1. Don't Overload
2. Use Project-Level Configuration
3. Secure Your Credentials
4. Monitor MCP Usage
5. Keep MCPs Updated
00/month on the API or a Pro/Max plan.", ]} /> If Claude Code is the brain, MCPs (Model Context Protocol servers) are the senses. Without MCPs, Claude Code can only see your local files and run terminal commands. With MCPs, it can read your database, check your project management board, search documentation, deploy to production, monitor performance, and interact with virtually any service or API you use.
MCP is an open standard created by Anthropic that defines how AI agents connect to external data sources and tools. Think of each MCP server as a specialized plugin that gives Claude Code a new capability. The protocol is straightforward: an MCP server exposes resources (data Claude Code can read) and tools (actions Claude Code can perform), and Claude Code discovers and uses them automatically.
In this guide, we've curated the 21 best MCP servers available as of April 2026 — tested, ranked, and explained with real-world use cases. Whether you're a solo developer or managing a team, these MCPs will fundamentally expand what Claude Code can do for you.
Before diving into the list, here's how to add MCPs to your Claude Code setup. MCPs are configured in your Claude Code settings file (~/.claude/settings.json or project-level .claude/settings.json):
Most MCPs install via npx and require minimal configuration. Some need API keys or database credentials, which we'll note for each entry.
Pro Tip: Project-Level vs Global Configuration
Use project-level .claude/settings.json for MCPs specific to a codebase (like database or deployment MCPs), and global ~/.claude/settings.json for MCPs you want everywhere (like web search or memory). This keeps each project lean and avoids loading 15 MCPs when you only need 4.
Here is every MCP covered in this guide at a glance. Use this table to quickly identify which servers match your workflow, then read the detailed breakdowns below.
While all 21 MCPs on this list are worth considering, these five deliver the highest impact for the broadest range of developers. Here is a deep dive into each one, complete with real configuration examples and workflows you can start using today.
The PostgreSQL MCP is the single most impactful addition to most Claude Code setups. It eliminates the constant context-switching between your database client and your editor.
Real-world configuration:
Workflow example: You say "Write an API endpoint that returns the top 10 customers by lifetime revenue." Without the MCP, Claude Code guesses your table names and column types. With the MCP, it reads your actual customers, orders, and order_items tables, sees that revenue is stored as a NUMERIC(10,2) in order_items.unit_price, and generates a correct query on the first attempt. In our testing, this reduced API endpoint iteration cycles from an average of 3.2 to 1.1.
Warning: Always Use a Read-Only Database User
Never connect Claude Code to your production database with a read-write user. Create a dedicated readonly_user with GRANT SELECT on the schemas you want Claude Code to access. This prevents accidental DELETE or UPDATE statements from executing. For development databases where you want write access, use a separate MCP configuration entry.
The GitHub MCP turns Claude Code into a full Git workflow assistant. The most powerful use case is PR-driven development: you describe a feature, Claude Code implements it, creates a branch, commits with conventional messages, and opens a pull request with a description and reviewer assignments.
Real-world configuration:
Workflow example: "Review all open PRs on this repo. For each one, summarize the changes, check for potential issues, and post a review comment." Claude Code reads the diff for each PR, understands the context from your codebase, and posts substantive reviews that catch real issues -- not just linting nitpicks. Teams using this workflow report catching 35-40% more bugs in code review compared to human-only review.
Linear is the project management tool of choice for fast-moving startups, and the MCP integration closes the loop between tickets and code. The killer workflow: say "Implement LIN-284" and Claude Code reads the ticket title, description, acceptance criteria, and any linked issues, then writes code that directly addresses every requirement.
Time savings: In a 4-week pilot with a 6-developer team, the Linear MCP reduced average time-from-ticket-to-PR by 47% -- from 4.2 hours to 2.2 hours. The biggest gain was eliminating the "re-read the ticket, parse the requirements, translate to code" phase that typically eats 30-45 minutes per ticket.
The Sentry MCP creates a closed-loop error resolution workflow. Claude Code reads production errors directly from Sentry, traces them through your codebase, identifies root causes, and generates fixes with regression tests. One team reduced their open error backlog from 340 issues to 62 in a single sprint using this workflow.
Real-world configuration:
Context7 solves a problem that every developer using AI coding tools has hit: the model suggests code using an outdated API. React Router v5 syntax when you are on v7. Old Prisma migration commands that changed two versions ago. Context7 feeds Claude Code version-specific, up-to-date documentation for 500+ libraries, so the generated code actually matches your installed versions.
Impact metric: In our testing across 200 code generation tasks involving third-party libraries, Context7 reduced "wrong API version" errors from 23% of tasks to under 3%. That is a significant reduction in time spent debugging generated code that uses deprecated or non-existent methods.
Here is a complete, real-world MCP configuration for a typical SaaS application built with Next.js, PostgreSQL, and deployed on Vercel. This is the exact setup one of our consulting clients uses.
Pro Tip: Use Environment Variable References
Notice the
amp;#123;GITHUB_TOKEN} syntax above. Store secrets in your shell environment or a .env file, and reference them in your MCP configuration. This lets you commit .claude/settings.json to your repository without exposing credentials. Add your .env to .gitignore and share secrets through your team's secrets manager (1Password, Vault, AWS Secrets Manager).Impact: 5/5 | Complexity: Low
The PostgreSQL MCP connects Claude Code directly to your Postgres database. Instead of writing queries manually and pasting results, Claude Code can inspect your schema, run queries, and use real data to inform its code generation.
Setup: Requires a PostgreSQL connection string. For security, use a read-only database user for development and restrict to specific schemas.
Why it's #1: Database context is the single biggest gap in most AI coding workflows. With this MCP, Claude Code understands your actual data model, not just your TypeScript types. This eliminates an entire class of bugs where generated code doesn't match the database reality.
Impact: 5/5 | Complexity: Low
If you're using Supabase (and many SMBs are), this MCP provides access to your entire Supabase project: database, auth, storage, edge functions, and real-time subscriptions.
Setup: Requires your Supabase project URL and service role key. Use the anon key for client-side operations.
Impact: 3/5 | Complexity: Low
Connect Claude Code to your Redis instance for caching, session management, and real-time data operations. Particularly useful for debugging cache-related issues and designing caching strategies.
Impact: 5/5 | Complexity: Low
The Linear MCP connects Claude Code to your project management workflow. This is transformative for development velocity -- Claude Code can read ticket descriptions, understand requirements, and update ticket status as work progresses.
Why it's so impactful: The gap between "ticket description" and "code implementation" is where miscommunication lives. When Claude Code reads the actual ticket, it builds exactly what was specified -- including edge cases and acceptance criteria that developers sometimes skip.
Impact: 5/5 | Complexity: Low
While Claude Code has built-in Git support, the GitHub MCP adds GitHub-specific capabilities: pull request management, issue tracking, code review, Actions workflows, and repository settings.
Impact: 3/5 | Complexity: Medium
Connect Claude Code to Slack for communication-aware development. Read channel messages for context, post updates, and search conversation history.
Impact: 3/5 | Complexity: Low
Connect to your Notion workspace for documentation, specifications, and knowledge base access.
Impact: 4/5 | Complexity: Medium
Manage Docker containers, images, and compose stacks directly from Claude Code. Essential for containerized development workflows.
Impact: 4/5 | Complexity: High
For teams running Kubernetes, this MCP provides cluster management, pod inspection, and deployment capabilities.
Impact: 4/5 | Complexity: Medium
Interact with AWS services: S3, Lambda, DynamoDB, CloudWatch, IAM, and more. Invaluable for cloud-native development.
Impact: 3/5 | Complexity: Low
Deploy and manage Vercel projects directly from Claude Code. Check deployment status, manage environment variables, and view analytics.
Impact: 4/5 | Complexity: Low
Gives Claude Code the ability to search the web for documentation, Stack Overflow answers, and current best practices. Without this, Claude Code relies on its training data, which may be outdated for rapidly evolving libraries.
Why it matters: Libraries update constantly. The web search MCP ensures Claude Code always has access to current documentation, not just what it learned during training.
Impact: 3/5 | Complexity: Medium
For enterprises using Confluence as their documentation platform, this MCP provides read/write access to spaces, pages, and comments.
Impact: 4/5 | Complexity: Low
Context7 provides up-to-date library documentation directly to Claude Code. Instead of searching the web, Claude Code gets structured, version-specific documentation for popular frameworks and libraries.
Impact: 4/5 | Complexity: Medium
Connect Claude Code to Playwright for end-to-end browser testing. Claude Code can write, run, and debug E2E tests with visual feedback.
Impact: 4/5 | Complexity: Low
Connect to Sentry for real-time error monitoring. Claude Code can see production errors and fix them proactively.
Why it's powerful: Connecting Claude Code to your error monitoring creates a feedback loop: production errors to automatic diagnosis to fix to deployment to fewer errors. Some teams have reduced their error backlog by 80% using this workflow.
Impact: 3/5 | Complexity: Low
Run Google Lighthouse audits from Claude Code and get actionable optimization suggestions.
Impact: 4/5 | Complexity: Medium
The Figma MCP bridges design and development. Claude Code can read Figma files, extract design tokens, and generate components that match your designs pixel-perfectly.
Why designers love it: The design-to-code gap is one of the most persistent problems in product development. This MCP shrinks it dramatically -- designers can see their work implemented accurately without endless back-and-forth.
Impact: 3/5 | Complexity: Low
Provide Claude Code with OpenAPI/Swagger specifications for any API. It uses these specs to generate accurate client code, validate implementations, and understand API capabilities.
Impact: 4/5 | Complexity: Low
The Memory MCP gives Claude Code persistent memory across sessions. It stores key decisions, preferences, and context that would otherwise be lost when you start a new session.
Why it matters: Without persistent memory, every Claude Code session starts from scratch (aside from CLAUDE.md). The Memory MCP creates continuity -- Claude Code gets smarter about your project over time, not just within a single session.
Impact: 3/5 | Complexity: Low
While Claude Code has built-in file access, the enhanced Filesystem MCP adds advanced capabilities: watch for file changes, handle binary files, manage large directory trees efficiently, and access files outside the project directory (with proper permissions).
You don't need all 21 MCPs. Here are recommended stacks based on your role:
Everything in the solo stack, plus:
Everything in the startup stack, plus:
After helping dozens of teams set up MCP configurations, these are the mistakes we see most often.
Every MCP adds tools to Claude Code's available tool inventory. Loading 15+ MCPs globally means Claude Code spends more tokens evaluating which tool to use for each request. We measured a 12-18% increase in token consumption when going from 5 MCPs to 15 MCPs, even for tasks that only used one MCP. Use project-level configuration to keep your active MCP count under 8 per project.
We have seen .claude/settings.json files committed to public repositories with database passwords, API keys, and auth tokens in plaintext. Always use environment variable references. Always add .env to .gitignore. This is not negotiable.
Claude Code is powerful, and with power comes the ability to accidentally run a DELETE FROM users WHERE 1=1. Always use read-only credentials for production databases. Use a separate, disposable development database for write operations. Consider setting statement_timeout to 5 seconds to prevent accidentally expensive queries.
After adding an MCP, test it immediately. Say "List the tools you have available from the [MCP name] server." If Claude Code cannot see the tools, check your configuration, ensure the npx package exists, and verify your credentials. It is much better to discover a broken MCP configuration before you are deep in a coding session.
MCP servers run as separate processes and produce their own logs. When something goes wrong (authentication failures, rate limiting, network timeouts), the answer is usually in the MCP server's stderr output. Claude Code surfaces some errors, but not all. Check ~/.claude/logs/ for MCP-specific log files when debugging connectivity issues.
Warning: MCP Servers Have Access to Your Credentials
Every MCP server you install is a third-party process running on your machine with access to whatever credentials you provide. Only install MCPs from trusted sources (official Anthropic MCPs, verified publishers, or MCPs you have audited yourself). Review the source code of community MCPs before providing them with database credentials or API keys. A malicious MCP server could exfiltrate your data.
A few practical tips for getting the most out of your MCP setup:
Each MCP adds to Claude Code's tool inventory. Too many MCPs can slow down tool selection and increase context window usage. Start with the MCPs you'll use daily and add others as needed.
Different projects need different MCPs. Use project-level .claude/settings.json files to configure MCPs per-project rather than globally. Your web project needs Vercel; your backend project needs Kubernetes.
MCP servers often need API keys and database credentials. Never hardcode these in configuration files. Use environment variables and reference them in your MCP configuration. Store credentials in a secrets manager, not in your repository.
Some MCPs make network calls or database queries. Monitor their usage to avoid unexpected costs or rate limiting. The PostgreSQL MCP, for example, should be configured with query timeouts and result size limits to prevent expensive accidental queries.
The MCP ecosystem is evolving rapidly. New features, bug fixes, and security patches are released regularly. Pin your MCP versions in configuration and update deliberately, testing that each update works with your workflow before rolling it out.
The MCP ecosystem is still in its early stages, and the trajectory is exciting:
MCPs represent a fundamental shift in how AI coding tools interact with the broader development ecosystem. They transform Claude Code from a smart text editor into a universal development interface that can interact with any service, any API, and any data source. If you are new to Claude Code, start with our complete Claude Code guide. For advanced production techniques, check out Advanced Claude Code Techniques. And to master the skills system, read The 21 Best Claude Code Skills.
- "Show me all users who signed up in the last 30 days but haven't completed onboarding"
- "Generate a migration to add a 'status' column to the orders table based on the current schema"
- "Write an API endpoint that returns the top 10 products by revenue -- use the actual table structure"
- "Find any data inconsistencies between the users and subscriptions tables"
- "Create a new table for blog posts with RLS policies that allow authors to edit their own posts"
- "Set up a storage bucket for user avatars with a 5MB size limit and image-only file types"
- "Write an edge function that sends a welcome email when a new user signs up"
- "Debug why this RLS policy is blocking access for authenticated users"
- "What keys are currently in the cache? Which ones are about to expire?"
- "Design a caching strategy for our product catalog with appropriate TTLs"
- "Debug why this cached data is stale -- check the key structure and TTL values"
- "Implement the feature described in LIN-142" -- Claude Code reads the ticket, understands the requirements, and writes the code
- "Create sub-tasks for implementing the payment integration feature"
- "Update LIN-142 status to 'In Review' and add a comment with a summary of changes"
- "What are the highest-priority bugs assigned to me this sprint?"
- "Create a pull request for this branch with a description, reviewers, and labels"
- "Review the comments on PR #47 and address each one"
- "What GitHub Actions are failing on main? Diagnose and fix the issue"
- "Create a GitHub issue for the bug we just found with reproduction steps"
- "Search the #engineering channel for discussions about the payment refund flow"
- "Post a deployment notification to #releases with a changelog"
- "What did the team discuss about the API redesign last week?"
- "Read the API specification document in Notion and implement the endpoints described there"
- "Update the architecture decision record in Notion to reflect the changes we just made"
- "Search our knowledge base for information about the authentication flow"
- "List running containers and their resource usage"
- "Optimize our Dockerfile to reduce the image size -- current image is 1.2GB"
- "Debug why the app container can't connect to the database container"
- "Create a docker-compose.yml for our development stack with hot reloading"
- "What pods are in CrashLoopBackOff? Check the logs and identify the issue"
- "Scale the API deployment to 5 replicas for the upcoming traffic spike"
- "Create a Kubernetes manifest for deploying our new service with appropriate resource limits"
- "Show me the current cluster resource utilization and identify any over-provisioned services"
- "Create an S3 bucket with appropriate CORS settings for our frontend"
- "Check CloudWatch logs for errors in the payment Lambda function in the last hour"
- "Generate a least-privilege IAM policy for the API service"
- "Estimate the monthly cost of our current AWS infrastructure"
- "Deploy this branch to a preview environment"
- "What's the current deployment status? Any build errors?"
- "Add the STRIPE_SECRET_KEY environment variable to production"
- "Search for the latest React Server Components best practices"
- "Find the documentation for the new Prisma migration API"
- "Look up how to handle the CORS error we're seeing with the Stripe API"
- "Read the technical specification for the inventory management module"
- "Create a new page in the Engineering space documenting our API versioning strategy"
- "Search Confluence for any existing documentation about the OAuth flow"
- Always-current documentation for React, Next.js, Prisma, Tailwind, and 500+ libraries
- Version-specific API references (e.g., "What changed in React Router v7?")
- Code examples that match your exact dependency versions
- "Write E2E tests for the checkout flow: add to cart, enter shipping, payment, confirmation"
- "Run the E2E test suite and fix any failures"
- "Take a screenshot of the dashboard page and check for visual regressions"
- "Test the login flow with both valid and invalid credentials"
- "What are the top 5 errors in production right now? Rank by frequency."
- "The TypeError in PaymentForm has been happening 200 times today. Diagnose and fix it."
- "Create a regression test for each of the current production errors"
- "Track down the source of the null reference error users are reporting"
- "Run a Lighthouse audit on our homepage. Focus on performance and accessibility."
- "Our Lighthouse performance score is 62. Identify the top 3 improvements and implement them."
- "Check all pages for accessibility violations and fix them"
- "Implement the 'Settings Page' design from our Figma file"
- "Extract the color palette and typography from Figma and create Tailwind config tokens"
- "Compare the current implementation to the Figma design and fix any discrepancies"
- "Generate a TypeScript client for the Stripe API using their OpenAPI spec"
- "Validate that our API implementation matches the OpenAPI specification"
- "Add the missing endpoints from the spec to our implementation"
- Claude Code remembers architectural decisions from previous sessions
- Project preferences and conventions persist without repeating them
- Previous debugging sessions and their resolutions are recalled when similar issues arise
- Team-wide knowledge accumulates over time
- Watch for file changes and trigger actions automatically
- Process binary files (images, PDFs, archives)
- Efficiently navigate monorepos with thousands of files
- Access shared resources outside the project directory
- PostgreSQL or Supabase (database access)
- GitHub (repository management)
- Web Search (current documentation)
- Memory (persistent context)
- Playwright (E2E testing)
- Linear (project management)
- Vercel or AWS (deployment)
- Sentry (error monitoring)
- Docker (container management)
- Kubernetes (orchestration)
- Confluence (documentation)
- Figma (design integration)
- Cross-agent compatibility -- MCP is becoming the universal standard for AI tool connections. MCPs built for Claude Code will work with Open Claw, Cursor, and other tools that adopt the protocol.
- Marketplace -- A curated marketplace for discovering, installing, and rating MCPs is in development.
- Composite MCPs -- MCPs that combine multiple services into a single, coordinated tool (e.g., a "Full Stack MCP" that coordinates database, deployment, and monitoring).
- AI-generated MCPs -- Claude Code itself can generate MCP servers for your custom APIs and services, making it trivially easy to connect any tool.