Source-reviewed developer tooling
Find developer-focused Model Context Protocol servers using evidence from official documentation, source repositories, and package registries.
Evidence reviewed from primary sources. No paid rankings or fabricated scores.
GitHub MCP Server
First-party
Connect coding agents to repositories, issues, pull requests, Actions, and GitHub security data through a configurable set of tools.
by GitHub
stdio · streamable-http
Go binary or Docker; hosted remote service also available
Context7
First-party
Retrieve version-aware library documentation and examples inside an agent conversation before code is generated.
by Upstash
streamable-http · stdio
Hosted service or Node.js 18+ via npm
Playwright MCP
First-party
Let agents navigate and operate web pages through Playwright's structured accessibility snapshots and persistent browser state.
by Microsoft
stdio · sse
Node.js 18+ via npm; Docker image also available
Chrome DevTools MCP
First-party
Give coding agents direct access to Chrome automation, console and network diagnostics, screenshots, and performance traces.
by Google Chrome DevTools
stdio
Node.js LTS, npm, and current Chrome or Chrome for Testing
Sentry MCP
First-party
Investigate Sentry issues, events, traces, releases, and project context from human-in-the-loop coding agents.
by Sentry
streamable-http · stdio
Hosted Cloudflare service or Node.js via npm for local stdio
Grafana MCP Server
First-party
Query observability data, inspect and manage dashboards and alerts, and work with Grafana incident tooling from MCP clients.
by Grafana Labs
stdio · sse · streamable-http
Go binary, uvx package, Docker, or Helm; hosted Grafana Cloud endpoint available
Supabase MCP Server
First-party
Let development agents inspect and manage Supabase databases, projects, logs, Edge Functions, branches, and documentation.
by Supabase
streamable-http · stdio
Supabase-hosted service; local development endpoint and npm package options also exist
Notion MCP
First-party
Search, read, create, and update Notion workspace content through Notion's hosted OAuth-based MCP service.
by Notion
streamable-http · sse
Notion-hosted remote service
Atlassian Rovo MCP Server
First-party
Bring Jira, Jira Service Management, Confluence, Bitbucket, and Rovo search into external AI clients through Atlassian's hosted server.
by Atlassian
streamable-http
Atlassian-hosted remote service
Azure DevOps MCP Server
First-party
Expose Azure Boards, Repos, Pipelines, test plans, wikis, search, and security context to development agents.
by Microsoft
streamable-http · stdio
Azure DevOps-hosted service or Node.js 20+ via npm
AWS Documentation MCP Server
First-party
Search, read, and follow related pages across the official AWS documentation set from an MCP client.
by AWS Labs
stdio
Python 3.10+ via uvx; an AWS-maintained container image is also available
Cloudflare API MCP Server
First-party
Explore and call the Cloudflare API through a compact, hosted MCP endpoint with OAuth or scoped API tokens.
by Cloudflare
streamable-http
Cloudflare-hosted remote service
MCP Toolbox for Databases
First-party
Connect agents to many database engines through prebuilt tools or a controlled YAML-defined tool layer.
by Google
stdio · streamable-http
Go binary or container; an npm launcher is available for convenience
Firecrawl MCP Server
First-party
Give an agent structured web search, scraping, crawling, extraction, interaction, and research tools through Firecrawl.
by Firecrawl
stdio · streamable-http
Node.js 22+ for the local npm package, or Firecrawl's hosted remote service
Figma MCP Server
First-party
Read design context, variables, assets, and Code Connect mappings—and, on supported workflows, create or edit native Figma content.
by Figma
streamable-http
Figma-hosted remote service; an optional local endpoint is built into the Figma desktop app
Codebase Memory MCP
Community
Index a repository into a local structural knowledge graph for fast symbol search, call tracing, architecture, and impact analysis.
by DeusData
stdio
Native static binary; npm requires Node.js 18+ and PyPI requires Python 3.8+ for distribution wrappers
Serena
Community
Add symbol-aware code retrieval, editing, refactoring, project memory, and optional IDE-backed operations to an MCP coding agent.
by Oraios AI
stdio · streamable-http
Python 3.11–3.14 managed with uv; the quick start selects Python 3.13
n8n-MCP
Community
Give agents searchable n8n node knowledge, workflow validation, and optional API-backed workflow management.
by czlonkowski
stdio · http
Node.js 18+ through npm, Docker, or a publisher-hosted service
Postgres MCP Pro
Community
Inspect PostgreSQL schemas, execute guarded SQL, explain plans, diagnose health, and test index recommendations.
by Crystal DBA
stdio · sse
Python 3.12+ through uvx or pipx; Docker image is recommended by the publisher
Linear MCP Server
First-party
Find and update Linear issues, projects, comments, initiatives, milestones, documents, and status updates through Linear's hosted MCP service.
by Linear
streamable-http
Linear-hosted remote service
Try removing a filter or using a broader capability term.
MCP server selection guide
The best MCP server is the smallest trusted integration that completes a real task with clear permissions, maintained code, and a transport your client supports. Use this directory to compare source-reviewed options, then verify the publisher, requested access, side effects, maintenance, and task fit before installation.
Discovery
No single discovery source provides both complete coverage and a security guarantee. A registry is designed for structured publication and lookup. A directory adds categories, comparisons, and editorial navigation. A marketplace usually adds an installation or commercial layer, while an awesome list is normally a community-maintained collection. The label does not tell you how deeply an entry was reviewed.
| Discovery source | Best use | Still verify |
|---|---|---|
| Official MCP Registry | Machine-readable discovery and publisher namespaces | Code quality, permissions, operational safety, and task fit |
| Graphify MCP directory | Human-readable comparison by capability, client, runtime, and transport | Current package instructions, credentials, and your threat model |
| Vendor documentation | Finding a first-party server or hosted endpoint | OAuth scopes, account boundary, data handling, and cost |
| Package registry | Inspecting versions and installing npm, PyPI, container, or other packages | Repository ownership, install scripts, dependencies, and release provenance |
| Client marketplace or awesome list | Compatibility-led browsing and broad community discovery | Publisher trust, requested access, maintenance, and side effects |
Registry inclusion is a useful discovery signal, not a production-readiness badge. Likewise, “open source” means the implementation can be inspected; it does not prove that the reviewed version is maintained, dependency-safe, or deployed with least privilege.
Selection by task
Start with the system the agent must reach and the exact action it needs to perform. Product profiles below remain the canonical source for current setup details; this table helps narrow the directory without pretending that one server wins every job.
GitHub MCP Server and Azure DevOps MCP connect repositories, issues, pull requests, and CI context.
Boundary: write tools can publish code or alter project state.Context7, Serena, and Codebase Memory MCP cover current docs, symbols, and repository memory.
Boundary: provenance, index freshness, and repository scope determine reliability.Playwright MCP, Chrome DevTools MCP, and Firecrawl MCP support testing, inspection, and crawling.
Boundary: browser sessions and extracted pages may contain credentials or untrusted instructions.Supabase MCP, Postgres MCP, and Database Toolbox expose schemas and queries.
Boundary: separate read-only exploration from writes, migrations, and administration.Atlassian MCP, Linear MCP, and Notion MCP connect tickets, plans, and team knowledge.
Boundary: record creation and status changes have organizational impact.n8n MCP, Sentry MCP, and Grafana MCP connect workflows and production evidence.
Boundary: one action can trigger downstream writes or expose sensitive telemetry.Search and content operations
“Search MCP” is not one product category. Web discovery, documentation retrieval, repository search, and enterprise knowledge search use different indexes and permission models. Start with the corpus that contains the answer, then require the server to return enough provenance—a source URL, document identifier, repository path, or version—to verify what the model says.
| Search target | Useful server shape | Evaluation question |
|---|---|---|
| Public web | Search-provider, crawler, or browser server | Does the result retain stable URLs, retrieval time, and enough page context to cite? |
| Library documentation | Version-aware documentation server such as Context7 | Can the client select the correct library and version, and is the upstream source identifiable? |
| Source code | Repository, symbol, or code-index server | Which branch and commit were indexed, and can results point back to files and symbols? |
| Internal knowledge | Vendor-supported connector that preserves workspace permissions | Are tenant isolation and document-level access controls enforced before retrieval? |
SEO work normally combines several evidence systems rather than relying on a single “SEO MCP server”: search-performance and analytics data, licensed keyword or SERP data, crawling and rendered-page inspection, content/entity analysis, and an issue tracker for implementation. Keep the reporting path read-only. Publishing, redirect changes, index controls, and production CMS edits should use separate write tools with explicit confirmation.
Retrieve a bounded result set, preserve provenance, distinguish quoted evidence from model inference, and validate important claims against a primary source. Treat page text and retrieved documents as untrusted input: search access should not silently authorize the same agent to publish content, change production settings, or transmit private data.
Terminology
Search results frequently mix protocol roles with products. Separating them prevents a directory visitor from trying to install a tool schema as a server or assuming that an SDK is a ready-to-run integration.
Installable or hosted integrations that expose capabilities to clients. Entries should identify the publisher, package or endpoint, transport, authentication, and effects.
The callable operations exposed by a connected server. A client discovers them at runtime, commonly through tools/list; the result can vary by version, role, and feature flag.
Applications that connect to servers. Compare support for local and remote transports, authentication, approvals, resources, prompts, and organization policy.
Developer packages used to implement clients and servers. A library can simplify protocol messages and transports, but it is not itself a production server or security policy.
Record the live capability list during evaluation rather than copying a README table. Hosts can hide tools, servers can conditionally expose them, and authentication may alter what one user sees. Review tool descriptions and input schemas before allowing unattended actions, especially when two servers publish similar tool names.
Evaluation method
A public endpoint can be first-party or anonymous, and an open repository can still be inactive or configured with excessive permissions. Move a candidate into a working toolset only after it passes all five checks.
Build ecosystem
An MCP SDK or framework helps developers implement protocol messages, capability negotiation, transports, and typed tool definitions. A language-specific MCP server is a runnable integration built with those primitives. Directories should keep these categories separate because their evaluation criteria differ: users judge a server by the system it reaches and the permissions it requests, while maintainers judge an SDK by specification compatibility, API stability, transport support, examples, and release cadence.
Start with the runtime already used by the service and the SDK maintained for the protocol features you need. Verify the exact repository and package before installing TypeScript, Python, C#, Java, Kotlin, Swift, Rust, Go, or other ecosystem packages; similarly named community libraries can have different coverage and release policies. Check support for your transport, authentication approach, structured outputs, cancellation, logging, and testing rather than choosing by download count alone.
Queries such as React MCP, Svelte MCP server, Rails MCP server, C# MCP, npm MCP, or PyPI MCP may describe an official integration, a community project, an SDK example, or a server generator. Confirm the entity before comparing it. A reference server can demonstrate the protocol without being hardened for production credentials, tenancy, rate limits, audit logs, or destructive actions.
Freshness
MCP changes across the specification, SDKs, registries, clients, and individual servers. General news is useful for discovery, but an implementation decision should trace back to a primary release note, tagged source release, or versioned documentation page. A directory review date tells you when evidence was checked; it does not guarantee that a package or hosted endpoint stayed unchanged afterward.
Review the versioned MCP specification for lifecycle, capability, and transport changes.
Follow official project releases for publication metadata, package APIs, compatibility notes, and deprecations.
Confirm which transports, authentication flows, approval controls, and protocol features the application actually enables.
Review endpoint, package, tool, scope, pricing, and security-advisory changes at the source that operates the integration.
Re-review immediately when a server changes publisher, endpoint, authentication method, transport, or write surface. For routine maintenance, assign an owner and a review interval based on impact: production and write-capable integrations deserve a shorter cycle than a read-only documentation server in an isolated profile.
Operations
Teams should maintain a small approved catalog even when discovery starts in a public registry or marketplace. An internal record gives every integration an owner and makes changes reviewable instead of relying on a copied configuration that no one maintains.
| Catalog field | Why it matters |
|---|---|
| Canonical name, endpoint, and package | Prevents lookalike installation and records exactly what runs. |
| Publisher, repository, and internal owner | Establishes provenance and accountability. |
| Supported clients, transport, and approved version | Prevents configuration drift and makes upgrades reviewable. |
| Credentials, scopes, and read/write effects | Supports least privilege and approval policy. |
| Data classification and network boundary | Prevents sensitive information from crossing an unapproved service. |
| Rollback, revocation, and last review date | Shortens incident response and exposes stale approvals. |
During testing, record the live tool list after initialization. Available tools can change with the server version, authenticated role, and feature flags, so a directory screenshot is not a substitute for inspecting the connected server.
FAQ
An MCP server list is a collection of server records. It may be a registry API, an editorial directory, a client marketplace, a package index, or a community-maintained list. Check how entries are submitted and reviewed before treating a list as a trust signal.
Not necessarily. A public endpoint may require a paid account, usage fees, or OAuth access to another service. Public describes availability, not price or permission scope.
Open source makes inspection possible, but safety still depends on the maintainer, version, dependencies, configuration, credentials, and deployment boundary. Review the exact code and release you plan to run.
Match the transports and authentication methods supported by the client to the server installation guide. Then connect it in a test profile and inspect the live capabilities before granting write access.
Use awesome lists for discovery, not bulk installation. Install individual servers only after verifying the publisher, permissions, transport, maintenance, and fit for a defined task.