The Model Context Protocol (MCP) gives AI agents a standardized way to interact with external tools, databases, APIs, and file systems. Instead of building custom integrations for every service, developers can install an MCP server and hand their agent structured access to real-world capabilities.
The ecosystem has kept expanding. Anthropic's official registry, vendor-maintained repositories, and independent developers have collectively published thousands of MCP servers. Some are polished and production-ready. Others are weekend projects that break on install. The signal-to-noise ratio is still a real problem.
This guide cuts through the noise. We evaluated MCP servers across eight categories based on documentation quality, maintenance activity, capability coverage, and security posture. Here are the ones actually worth integrating into your agent workflows, current as of September 2026.
One structural change worth flagging up front: a number of vendors that originally shipped their integration as a community reference implementation under the `@modelcontextprotocol` namespace have since taken ownership of it directly — publishing their own repo, and in several cases (GitHub, Atlassian, Notion, Slack) offering an official remote server you connect to over OAuth instead of running locally with a stored API key. Where that's happened, we've noted it below. Package names and install paths shift; treat the commands in this guide as a starting point and confirm the current one on the vendor's own docs before you run it.
How We Evaluated
Every MCP server listed here was assessed on four criteria:
- Maintenance status — Active commits within the last 90 days, responsive issue tracking
- Documentation — Clear setup instructions, capability descriptions, and configuration examples
- Security model — Explicit permission scoping, credential handling, sandboxing or OAuth options
- Practical utility — Does it solve a real workflow problem, or is it a demo?
Developer Tools
GitHub MCP Server
The GitHub MCP server is the most widely adopted server in the ecosystem. It exposes repository management, issue tracking, pull request workflows, code search, and file operations through MCP's tool interface.
Key capabilities:
- Create, read, update issues and pull requests
- Search code across repositories
- Manage branches, commits, and file contents
- Access GitHub Actions workflow status
- Fork repositories and manage collaborators
Install:
npx @modelcontextprotocol/server-githubNote: GitHub now maintains its own official server directly rather than through the original community reference implementation, and offers it both as a local binary and as a hosted remote server with OAuth. If you're setting this up fresh, start from GitHub's own MCP documentation rather than the legacy `npx` package.
Configuration requires: A GitHub personal access token with appropriate scopes (or, for the remote server, an OAuth login). Fine-grained tokens are recommended over classic tokens for tighter permission control.
Gotchas: Rate limiting applies based on your GitHub plan. The server does not cache responses, so agents making rapid successive calls can exhaust API limits quickly. Scope your token narrowly — an agent with delete permissions on repositories is a risk you don't need.
GitLab MCP Server
For teams on GitLab, this server mirrors much of the GitHub server's functionality: merge request management, issue tracking, pipeline status, and repository operations.
Install:
npx @modelcontextprotocol/server-gitlabGotchas: Self-hosted GitLab instances require additional configuration for API endpoints. Some enterprise features (epics, security dashboards) have limited coverage.
Linear MCP Server
Linear's MCP server provides access to issue tracking, project management, and team workflows. It is particularly useful for agents that triage bugs, create issues from error logs, or update project status automatically.
Key capabilities:
- Create and update issues with labels, priorities, and assignments
- Query project and cycle status
- Search across workspace issues
- Manage team workflows and states
Install:
npx @modelcontextprotocol/server-linearGotchas: Requires a Linear API key. Write operations should be tested in a sandbox workspace first, as agents can create issues at high volume if unconstrained.
Jira and Confluence MCP
Atlassian has since shipped an official remote MCP server covering both Jira and Confluence, authenticated via OAuth rather than a stored token — worth using in place of the older community-maintained local packages if your workspace is on Atlassian Cloud. Community servers still exist for Data Center deployments and for teams not yet on the remote path; the most actively maintained ones provide issue CRUD, JQL search, sprint management, and board operations.
Gotchas: Confirm whether your Confluence/Jira deployment is Cloud or Data Center before picking a server — the auth model and available endpoints differ. Community server quality still varies; check commit history before adopting one.
Databases
PostgreSQL MCP Server
This server gives agents direct SQL access to PostgreSQL databases. It supports read and write operations, schema inspection, and query execution with parameterized inputs.
Key capabilities:
- Execute arbitrary SQL queries (SELECT, INSERT, UPDATE, DELETE)
- Inspect database schema — tables, columns, types, constraints
- List databases and tables
- Run parameterized queries to prevent SQL injection
Install:
npx @modelcontextprotocol/server-postgresGotchas: This server can execute destructive queries. In production, connect through a read-only replica or a database user with restricted permissions. Never give an agent DROP or TRUNCATE privileges on production data.
SQLite MCP Server
A lightweight option for local development and prototyping. The SQLite server operates on local database files and supports full SQL operations.
Install:
npx @modelcontextprotocol/server-sqliteGotchas: File path permissions matter. The server needs read/write access to the database file and its directory (for WAL mode). Not suitable for concurrent access scenarios.
Supabase MCP Server
Supabase's official MCP server provides access to their PostgreSQL database, authentication, storage, and edge functions through a unified interface.
Key capabilities:
- Database queries via Supabase's client library
- User management and authentication operations
- File upload and retrieval from Supabase Storage
- Edge function invocation
Gotchas: Requires a Supabase project URL and service role key. The service role key bypasses Row Level Security, so use the anon key where possible and apply RLS policies.
Cloud Infrastructure
AWS MCP Servers
AWS provides a suite of MCP servers covering their major services. The most useful ones cover S3, CloudFormation, CloudWatch, Lambda, and CDK operations.
Key capabilities:
- S3 bucket and object management
- CloudWatch log querying and metric retrieval
- Lambda function invocation and management
- CloudFormation stack operations
- CDK project scaffolding and deployment
Gotchas: AWS IAM permissions are complex. Create a dedicated IAM user or role for your MCP server with the minimum required permissions. Avoid using root credentials or overly broad policies like AdministratorAccess.
Cloudflare MCP Server
Cloudflare's MCP server covers Workers, KV, R2, D1, and DNS management, and Cloudflare was one of the earlier vendors to push remote (Workers-hosted) MCP as the default connection method rather than a local process. It's well-documented and actively maintained by Cloudflare's own team.
Key capabilities:
- Deploy and manage Workers scripts
- CRUD operations on KV namespaces and keys
- R2 bucket and object management
- D1 database queries
- DNS record management
Gotchas: Requires a Cloudflare API token scoped to specific zones and services. Workers deployment through an agent should be tested in a staging environment first.
Vercel MCP Server
Vercel's MCP server enables deployment management, project configuration, environment variable handling, and domain operations.
Key capabilities:
- Trigger and monitor deployments
- Manage environment variables across environments
- Query deployment logs and build output
- Domain and DNS configuration
Gotchas: Deployment triggers from agents can run up build minutes quickly. Implement safeguards — rate limiting or approval gates — before giving an agent autonomous deploy access.
Communication
Slack MCP Server
The Slack MCP server enables agents to read channels, send messages, search message history, and manage threads. Slack now offers this as an official remote server with OAuth alongside the original local package, which is one of the more practical servers for workflow automation either way.
Key capabilities:
- Read and send messages in channels and DMs
- Search message history across the workspace
- Manage threads and reactions
- List channels and members
- Upload files to conversations
Install:
npx @anthropic/mcp-server-slackGotchas: Requires a Slack Bot token with appropriate OAuth scopes (or workspace OAuth login for the remote server). The chat:write scope lets agents post messages — ensure your agent has guardrails to prevent spam or unintended messages in public channels.
Gmail MCP Server
The Gmail server provides email read, search, compose, and send capabilities through Google's API.
Key capabilities:
- Search emails with Gmail's query syntax
- Read email content and metadata
- Compose and send emails
- Manage labels and threads
- Handle drafts
Gotchas: Google OAuth setup is involved — you need a Google Cloud project, OAuth consent screen, and credentials. The server stores refresh tokens locally. Sending emails autonomously is high-risk; implement human-in-the-loop approval for outbound messages.
Discord MCP Server
Community-maintained Discord servers enable message reading, sending, and channel management for Discord bots integrated with AI agents.
Gotchas: Discord's rate limits are aggressive. Community servers vary in maturity — evaluate carefully before production use.
Knowledge & Documents
Notion MCP Server
Notion now publishes its own official MCP server, providing structured access to pages, databases, blocks, and search across a workspace. It's well-suited for agents that need to read documentation, update project trackers, or create structured content.
Key capabilities:
- Search across workspace pages and databases
- Read and create pages with rich block content
- Query and update Notion databases
- Manage page properties and relations
Gotchas: Notion's API has pagination limits and rate limits. Large workspaces with thousands of pages may require careful query scoping. The integration needs explicit page-level sharing permissions.
Confluence MCP Server
Confluence is covered by Atlassian's official remote MCP server (see the Jira note above) for Cloud workspaces; community servers still provide page search, content reading, and space navigation for Data Center deployments.
Gotchas: Atlassian Cloud vs. Data Center APIs differ. Confirm your server version matches your Confluence deployment type.
Google Drive MCP Server
The Google Drive server enables file search, content reading, and metadata access across a user's Drive.
Key capabilities:
- Search files by name, type, and content
- Read file contents (Docs, Sheets, and plain text files)
- Access file metadata and permissions
- Navigate folder structures
Gotchas: Google OAuth setup required. File content extraction works best for Google-native formats; binary files (PDFs, images) may require additional processing. Be mindful of which files are shared with the integration.
Web & Browser Automation
Playwright MCP Server
Playwright's official MCP server, maintained by Microsoft, is the current default recommendation for browser automation over MCP — multi-browser support (Chromium, Firefox, WebKit), network interception, and accessibility-tree snapshots that give an agent structured page state instead of raw screenshots to interpret.
Key capabilities:
- Cross-browser automation (Chromium, Firefox, WebKit)
- Network interception and mocking
- Mobile device emulation
- Accessibility tree snapshots
- Video recording of browser sessions
Gotchas: Requires browser binaries to be installed (npx playwright install). The full install can be several hundred megabytes. Select only the browsers you need.
Puppeteer MCP Server (legacy)
The original Puppeteer MCP server wrapped headless Chrome for web scraping, screenshots, and form interaction. The reference implementation has since been deprecated in the official servers repo in favor of Playwright's actively maintained server above. If you're starting a new project, use Playwright; only reach for a Puppeteer-based server if you have an existing Puppeteer automation codebase you're wrapping rather than rewriting.
Gotchas: Headless Chrome is resource-intensive. Running it in constrained environments (serverless, containers) may require configuration tuning. JavaScript execution in the browser context is a security surface — sanitize inputs carefully.
Fetch MCP Server
A simpler alternative to full browser automation, the Fetch server makes HTTP requests and returns responses. Useful for API calls, webhook testing, and content retrieval from pages that don't require JavaScript rendering.
Install:
npx @modelcontextprotocol/server-fetchGotchas: No JavaScript rendering — dynamic single-page apps will return empty content. For those, use Playwright instead.
AI/ML Tools
Hugging Face MCP Server
The Hugging Face MCP server provides access to model search, dataset exploration, Space interaction, and paper search across the Hugging Face Hub.
Key capabilities:
- Search and query model repositories
- Explore datasets and their metadata
- Interact with Hugging Face Spaces
- Search academic papers
- Access documentation
Gotchas: Some operations require a Hugging Face API token. Model inference through the server can incur costs on paid inference endpoints.
Weights & Biases MCP Server
Community-maintained servers for W&B provide experiment tracking queries, run comparisons, and artifact management.
Gotchas: Community-maintained with variable update frequency. Verify compatibility with your W&B workspace version before relying on it for production workflows.
File Systems
Filesystem MCP Server
The filesystem server provides controlled access to local files and directories. It is the foundation for agents that need to read configuration files, write output, or manage local project structures.
Key capabilities:
- Read and write files with content type detection
- Create, move, and delete files and directories
- Search for files by name or pattern
- Get file metadata (size, modified time, permissions)
Install:
npx @modelcontextprotocol/server-filesystem /path/to/allowed/directoryGotchas: The directory argument is critical — it constrains the server to a specific path. Without it, the agent gets access to your entire filesystem. Always specify the narrowest directory scope possible.
S3 (via AWS MCP Server)
S3 access is provided through the AWS MCP server suite. It supports bucket listing, object CRUD, and presigned URL generation.
Gotchas: S3 operations can incur costs. Implement object size limits and request throttling for agents that process large volumes of files.
Comparison Table
| Server | Category | Maintainer | Auth Method | Read/Write | Maturity |
|---|---|---|---|---|---|
| GitHub | Dev Tools | GitHub (official) | PAT / Fine-grained token, or OAuth (remote) | Both | Stable |
| GitLab | Dev Tools | Anthropic (reference) | PAT | Both | Stable |
| Linear | Dev Tools | Anthropic (reference) | API Key | Both | Stable |
| Jira / Confluence | Dev Tools / Knowledge | Atlassian (official) | OAuth 2.0 (remote) / PAT (Data Center) | Both | Stable |
| PostgreSQL | Database | Anthropic (reference) | Connection string | Both | Stable |
| SQLite | Database | Anthropic (reference) | File path | Both | Stable |
| Supabase | Database | Supabase | Project URL + Key | Both | Stable |
| AWS (S3, Lambda, etc.) | Cloud | AWS | IAM credentials | Both | Stable |
| Cloudflare | Cloud | Cloudflare | API Token / OAuth (remote) | Both | Stable |
| Vercel | Cloud | Vercel | API Token | Both | Stable |
| Slack | Communication | Slack (official) | Bot Token / OAuth (remote) | Both | Stable |
| Gmail | Communication | Community | Google OAuth 2.0 | Both | Beta |
| Discord | Communication | Community | Bot Token | Both | Beta |
| Notion | Knowledge | Notion (official) | Integration Token / OAuth | Both | Stable |
| Google Drive | Knowledge | Community | Google OAuth 2.0 | Read-heavy | Beta |
| Playwright | Web | Microsoft | None | Both | Stable |
| Puppeteer | Web | Legacy reference | None | Both | Deprecated — use Playwright |
| Fetch | Web | Anthropic (reference) | None | Read | Stable |
| Hugging Face | AI/ML | Hugging Face | API Token (optional) | Read-heavy | Stable |
| W&B | AI/ML | Community | API Key | Read-heavy | Beta |
| Filesystem | File System | Anthropic (reference) | Directory path | Both | Stable |
Security Considerations
MCP servers are powerful precisely because they give AI agents access to real systems. That power requires careful security practices.
Permission Scoping
Every MCP server should be configured with the minimum permissions required for its task. This means:
- Database servers: Use read-only database users unless writes are explicitly needed. Never connect as a superuser.
- API tokens: Use fine-grained or scoped tokens. GitHub fine-grained PATs, AWS IAM policies with specific resource ARNs, Cloudflare tokens scoped to specific zones.
- Filesystem servers: Always specify the narrowest directory scope. A filesystem server with access to
/is an open door.
Remote vs. Local: A Real Security Tradeoff
Where a vendor offers an official remote (OAuth) MCP server — GitHub, Atlassian, Notion, Slack, and Cloudflare all do now — it typically beats the local alternative on security grounds: no long-lived API key sitting in a config file or environment variable on your machine, and the vendor controls token scope and rotation. The tradeoff is less control over exactly what runs and where, and you're trusting the vendor's hosted infrastructure rather than a process you can inspect. For anything touching production data, weigh that tradeoff deliberately rather than defaulting to whichever option was easier to set up.
Supply Chain Risks
Community-maintained MCP servers are npm packages or Python packages installed from public registries. The standard supply chain risks apply:
- Pin specific versions in your configuration rather than using
latest - Audit the source code before first installation — especially for servers that handle credentials
- Monitor for package ownership transfers (typosquatting and maintainer account compromise are real threats)
- Prefer servers published by the service vendor (GitHub, Cloudflare, Vercel, Notion, Atlassian) or by Anthropic's official reference repo over unfamiliar community packages
Sandboxing
For high-risk servers (filesystem, database, browser automation), consider running them in sandboxed environments:
- Docker containers with restricted network access and mounted volumes
- Virtual machines for full isolation
- Serverless functions with execution time limits and memory caps
Credential Management
MCP server credentials (API keys, tokens, connection strings) should be managed through environment variables, not hardcoded in configuration files. Use a secrets manager (AWS Secrets Manager, HashiCorp Vault, or .env files excluded from version control) to store and rotate credentials.
The Bottom Line: Your Essential Starter Kit
If you are building AI agent workflows and want to start with MCP servers that deliver immediate value, here is the recommended starter kit:
- Filesystem — Foundation for any local file operations. Scope it tightly.
- GitHub — If your code lives on GitHub, this server is essential for agent-driven development workflows. Use the official server; check whether the remote OAuth option fits your setup before defaulting to a local token.
- PostgreSQL or SQLite — Database access for agents that need to query or update structured data.
- Fetch — Lightweight HTTP access for API calls and content retrieval without the overhead of a full browser.
- Slack — If your team communicates on Slack, this enables agents to participate in workflows where humans already collaborate.
Start with these five. Add specialized servers (cloud infrastructure, browser automation, knowledge tools) as your agent workflows mature and the use cases justify the added complexity.
The MCP ecosystem is still moving fast — package names, hosting options, and which servers are considered "official" versus "reference" have all shifted since this space first took off, and they'll likely shift again. Evaluate each server against your specific needs, apply strict security practices, and avoid installing servers just because they exist. Always verify current install instructions and maturity status on the vendor's own docs before adopting one.
For building agents that use MCP servers effectively, tools like Claude and Cursor provide native MCP integration that makes configuration and testing straightforward.
FAQ
New reviews, every week.
One email when we publish. No hype, no spam, unsubscribe anytime.
Related reads
More from WildRun Reviews
Part of the WildRun AI network.