sunnamed434/awesome-mcp-registry：由 DeepSeek-V4.1-Flash 每周自策展的 MCP 服务器目录
sunnamed434/awesome-mcp-registry
sunnamed434/awesome-mcp-registry 是一个由 AI 自策展的 Model Context Protocol 服务器目录，从 GitHub 和官方 MCP Registry 发现服务器，每周由 DeepSeek-V4.1-Flash 分析并用公开可复现的信任公式打分。
A self-curating directory of Model Context Protocol servers — a Continuous AI experiment. Discovered from GitHub and the Official MCP Registry, analyzed by DeepSeek-V4.1-Flash weekly and scored with a published, reproducible trust formula.
Our bet: curation is a job for AI, not gatekeepers. No maintainers deciding what's "in", no PR queues, no politics — just a Continuous AI workflow that discovers, judges, and re-judges every server on merit, week after week. This list is a small proof of a bigger idea: that AI can own a real, useful, self-maintaining system end to end. Humans set the rules once; the AI runs it.
Using an AI agent on this repo? Point it at AGENTS.md first: this list is machine-generated, servers are nominated through an issue form (never a pull request), and PRs editing the generated files are closed automatically. Machine-readable map of the repo: llms.txt.
Trending This Week
| Server | Stars | Δ 7 days |
|---|---|---|
| mksglu/context-mode | 25368 | +1243 |
| mobile-next/mobile-mcp | 8641 | +899 |
| apify/apify-mcp-server | 9585 | +883 |
| DeusData/codebase-memory-mcp | 45785 | +758 |
| CursorTouch/Windows-MCP | 7674 | +336 |
Databases (3)
| Server | Trust | Stars | Description |
|---|---|---|---|
| googleapis/mcp-toolbox | 96/100 | 16564 (+70/wk) | An open source MCP server that connects AI agents and IDEs to enterprise databases like PostgreSQL, MySQL, BigQuery, and MongoDB through prebuilt and custom tools. |
| benborla/mcp-server-mysql | 79/100 | 2141 (+2/wk) | An MCP server that gives LLMs read-only (optionally write-enabled) access to MySQL databases for schema inspection and querying. |
| qdrant/mcp-server-qdrant | 73/100 | 1542 (+3/wk) | An official MCP server that stores and retrieves semantic memories in a Qdrant vector database via qdrant-store and qdrant-find tools. |
Dev Tools (24)
| Server | Trust | Stars | Description |
|---|---|---|---|
| github/github-mcp-server | 100/100 | 33359 (+126/wk) | Official GitHub MCP server that lets AI agents read repositories, manage issues and pull requests, and inspect CI/CD and security findings on GitHub. |
| microsoft/playwright-mcp | 99/100 | 37812 (+195/wk) | An MCP server that gives LLM clients browser automation capabilities through Playwright using structured accessibility snapshots. |
| czlonkowski/n8n-mcp | 95/100 | 23036 (+28/wk) | An MCP server that exposes n8n node documentation, properties, operations, and workflow templates to AI assistants so they can build n8n workflows. |
| oraios/serena | 92/100 | 29983 (+142/wk) | An MCP server that gives coding agents IDE-style semantic code retrieval, editing, refactoring, and debugging tools at the symbol level. |
| getsentry/MobileBuildMCP | 92/100 | 6458 (+24/wk) | An MCP server and CLI exposing Xcode/iOS-macOS build and project tools to AI coding agents. |
| mobile-next/mobile-mcp | 91/100 | 8641 (+899/wk) | An MCP server that lets agents automate and scrape native iOS and Android apps on simulators, emulators, and real devices through accessibility snapshots and coordinate-based interactions. |
| GLips/Figma-Context-MCP | 88/100 | 15950 (+34/wk) | An MCP server that fetches and simplifies Figma file layout and styling data for AI coding agents like Cursor. |
| mrexodia/ida-pro-mcp | 88/100 | 12466 (+107/wk) | An MCP server that bridges IDA Pro's disassembler/decompiler to LLM clients so an agent can inspect and annotate binaries during reverse engineering. |
| ChromeDevTools/chrome-devtools-mcp | 85/100 ⚠ | 52950 (+290/wk) | An MCP server that lets coding agents control and inspect a live Chrome browser via Puppeteer and Chrome DevTools for automation, debugging, and performance analysis. |
| yusufkaraaslan/Skill_Seekers | 84/100 | 15097 (+59/wk) | A Python tool that converts documentation sites, GitHub repos, PDFs and other sources into structured knowledge assets and Claude AI skills, with an advertised MCP integration exposing ~40 tools. |
| neka-nat/freecad-mcp | 84/100 | 2664 (+153/wk) | MCP server that lets MCP clients drive FreeCAD to create and edit CAD models, run Python scripts, inspect documents, and run FEM analyses. |
| mksglu/context-mode | 83/100 | 25368 (+1243/wk) | A context-window optimization tool for AI coding agents that sandboxes tool output and persists session memory, exposing this via an MCP server and hooks. |
| mixelpixx/KiCAD-MCP-Server | 83/100 | 2567 (+84/wk) | An MCP server that lets AI assistants drive KiCAD for PCB design automation, including schematic editing, component placement, routing, design-rule checks, and manufacturing exports. |
| Jpisnice/shadcn-ui-mcp-server | 82/100 | 3021 (+17/wk) | An MCP server that gives AI assistants access to shadcn/ui v4 component source, demos, blocks, and metadata across React, Svelte, Vue, and React Native. |
| DeusData/codebase-memory-mcp | 77/100 ⚠ | 45785 (+758/wk) | A native C MCP server that indexes codebases into a persistent knowledge graph and answers structural code queries for AI coding agents. |
| duty1g/x64dbg-mcp-server | 77/100 ⚠ | 2181 (+73/wk) | A native x64dbg plugin that runs an MCP server over HTTP, letting MCP-compatible AI assistants control the debugger (breakpoints, stepping, memory, registers, and more). |
| bgauryy/octocode | 77/100 | 945 (+1/wk) | Octocode is an evidence-first code research MCP server and CLI that lets agents search, read, and analyze local code and GitHub repositories with ripgrep, AST, and LSP tools. |
| robotmcp/ros-mcp-server | 75/100 | 1483 (+7/wk) | An MCP server that connects LLM clients to ROS/ROS2 robots through rosbridge, letting models inspect and control topics, services, actions, and parameters. |
| wonderwhy-er/DesktopCommanderMCP | 74/100 ⚠ | 9912 (+125/wk) | An MCP server that gives AI clients terminal command execution, filesystem search, and diff-based file editing capabilities. |
| CursorTouch/Windows-MCP | 71/100 ⚠ | 7674 (+336/wk) | An MCP server that lets AI agents control the Windows desktop through UI automation, app launching, window management, and simulated input. |
...and 4 more. See known_servers.json for the full list.
Cloud (1)
| Server | Trust | Stars | Description |
|---|---|---|---|
| containers/kubernetes-mcp-server | 91/100 | 2142 (+17/wk) | An MCP server that exposes Kubernetes and OpenShift cluster operations (pods, namespaces, events, Helm, Tekton) as tools for MCP clients. |
Productivity (6)
| Server | Trust | Stars | Description |
|---|---|---|---|
| sooperset/mcp-atlassian | 97/100 | 5966 (+21/wk) | An MCP server exposing Jira and Confluence tools (issue search/creation/updates, page search) to MCP clients like Claude Desktop and Cursor. |
| jacob-bd/gemini-notebook-mcp-cli | 86/100 | 6225 (+56/wk) | An MCP server (and CLI) providing programmatic access to Google Gemini Notebook/NotebookLM notebooks. |
| haris-musa/excel-mcp-server | 82/100 | 4212 (+10/wk) | An MCP server that lets AI agents create, read, and modify Excel workbooks (sheets, formulas, formatting, charts, pivot tables) without Microsoft Excel installed. |
| zcaceres/markdownify-mcp | 74/100 | 2999 (+1/wk) | An MCP server that converts PDFs, images, audio, DOCX/XLSX/PPTX, web pages, YouTube transcripts, and Bing results into Markdown. |
| stonianua/neither-mcp | 61/100 | 0 | Neither is a decision-memory MCP server for Cursor, Claude Desktop, and Claude Code that connects to the Neither workspace API via a workspace key. |
| worklittle/jobs-mcp | 59/100 | 0 | Public connector and skill files pointing to Worklittle's remote MCP server for searching over 4 million jobs with filters and saving jobs to a Worklittle account. |
Communication (2)
| Server | Trust | Stars | Description |
|---|---|---|---|
| Autoposting-ai/autoposting-mcp | 56/100 | 0 | Remote MCP server for Autoposting that lets MCP clients schedule, generate, and publish social media posts across X, LinkedIn, Instagram, Threads, and YouTube. |
| snow884/adam-network | 56/100 | 0 | An agent-oriented social network/messaging board that includes a bundled MCP server for AI assistants to read and post messages. |
Web Scraping (9)
| Server | Trust | Stars | Description |
|---|---|---|---|
| apify/apify-mcp-server | 95/100 | 9585 (+883/wk) | Apify's MCP server lets AI agents discover and run thousands of Apify Store scrapers and automation Actors to extract data from websites, social media, search engines, and e-commerce platforms. |
| firecrawl/firecrawl-mcp-server | 90/100 | 7551 (+30/wk) | Official Firecrawl MCP server that gives MCP clients web search, scraping, crawling, mapping, and page-interaction tools backed by the Firecrawl API. |
| brightdata/brightdata-mcp | 87/100 | 2661 (+2/wk) | Bright Data's MCP server exposes web search, page scraping, structured data extraction, and remote browser automation tools to AI agents over MCP. |
| xpzouying/xiaohongshu-mcp | 81/100 ⚠ | 16108 (+103/wk) | An MCP server that lets AI assistants log into and interact with xiaohongshu.com, including checking login status and publishing image-text posts. |
| Xquik-dev/x-twitter-scraper | 74/100 ⚠ | 210 (+4/wk) | A hosted X (Twitter) scraping API that also exposes an MCP integration for searching tweets, profiles, followers and replies. |
| feder-cr/invisible_playwright_mcp | 68/100 ⚠ | 31773 (+95/wk) | An MCP server that drives an anti-detect Playwright browser so coding agents can perform undetected web browsing, scraping, and computer-use tasks. |
| executeautomation/mcp-playwright | 64/100 | 5662 (+4/wk) | An MCP server that gives LLM clients Playwright-based browser automation, including navigation, screenshots, device emulation, scraping, and JavaScript execution. |
| epiral/bb-browser | 60/100 ⚠ | 6231 | A CLI and MCP server that lets AI agents drive the user's own logged-in Chrome to run site-specific commands (search, transcripts, stock quotes, job search) across 36 platforms. |
| hangwin/mcp-chrome | 55/100 | 12463 (+7/wk) | A Chrome extension plus local bridge that exposes the user's own browser as an MCP server for AI-driven automation, content analysis, and semantic tab search. |
AI & ML (1)
| Server | Trust | Stars | Description |
|---|---|---|---|
| BeehiveInnovations/pal-mcp-server | 66/100 | 11768 (+11/wk) | An MCP server that lets a single AI CLI orchestrate multiple model providers (Gemini, OpenAI, Anthropic, Grok, Ollama, etc.) and spawn external CLI subagents through tools like consensus and clink. |
Finance (1)
| Server | Trust | Stars | Description |
|---|---|---|---|
| vbkotecha/agentservices-api | 68/100 | 2 (+1/wk) | Hosted MCP server exposing paid x402/USDC crypto, market-intelligence, LLM, image, and TTS APIs to AI agents. |
Security (5)
| Server | Trust | Stars | Description |
|---|---|---|---|
| 0x4m4/hexstrike-ai | 73/100 | 12326 (+172/wk) | An MCP server that exposes 150+ cybersecurity and pentesting tools plus autonomous AI agents to LLM clients for automated security testing. |
| palisadeemail/palisade-mcp | 68/100 | 0 | A local stdio bridge that connects MCP clients to Palisade's remote email-authentication server for SPF, DKIM, DMARC, MTA-STS, and BIMI monitoring and management. |
| Orphograph/Orphograph | 68/100 | 0 | An MCP server that anchors file SHA-256 hashes to Bitcoin via OpenTimestamps and verifies the resulting receipts offline. |
| LaurieWired/GhidraMCP | 63/100 | 10273 (+70/wk) | An MCP server that bridges Ghidra's reverse-engineering capabilities (decompilation, symbol renaming, listing methods/imports/exports) to LLM clients. |
| doteyeso-ops/mcp-server-vibes-coded | 50/100 ⚠ | 3 | A Python MCP server exposing ~26 tools for agent supply-chain risk scanning, scanner consensus, reliability guards, attestation, and a paid x402-based agent commerce/memory catalog. |
Media (2)
| Server | Trust | Stars | Description |
|---|---|---|---|
| yuniko-software/minecraft-mcp-server | 58/100 | 770 (+10/wk) | An MCP server that lets AI assistants control a Minecraft character via the Mineflayer API to build, explore, and interact with the game world. |
| MiniMax-AI/MiniMax-MCP | 50/100 ⚠ | 1583 | Official MiniMax MCP server exposing text-to-speech, image generation, and video generation APIs to MCP clients. |
Search (3)
| Server | Trust | Stars | Description |
|---|---|---|---|
| perplexityai/modelcontextprotocol | 90/100 | 2551 (+7/wk) | Official Perplexity API Platform MCP server providing real-time web search, reasoning, and research tools to MCP clients. |
| blazickjp/arxiv-mcp-server | 84/100 | 3188 (+11/wk) | A local stdio MCP server that lets agents search arXiv, read original-LaTeX paper sections, extract BibTeX citations, and watch topics, keeping papers on disk. |
| BuyWhere/buywhere-mcp | 68/100 | 15 (+1/wk) | A TypeScript MCP server that lets AI agents search products, compare prices across Singapore/US/SEA retailers, and discover deals via the BuyWhere commerce API. |
Knowledge Base (3)
| Server | Trust | Stars | Description |
|---|---|---|---|
| coddingtonbear/obsidian-local-rest-api | 93/100 | 2986 (+20/wk) | An Obsidian plugin that exposes a secure local REST API and a built-in MCP server for reading, writing, searching, and patching notes in a vault. |
| upstash/context7 | 90/100 | 62668 (+198/wk) | Hosted MCP server that fetches current, version-specific documentation and code examples for libraries and injects them into LLM context. |
| datagouv/datagouv-mcp | 81/100 | 1599 (+2/wk) | Official MCP server for data.gouv.fr that lets AI chatbots search, explore, and analyze French national open datasets. |
How This Works
This registry is automatically maintained by a GitHub Actions workflow that runs weekly:
- Discover — searches GitHub and the Official MCP Registry for new servers; community nominations join the same queue
- Analyze — each repo is evaluated by AI (DeepSeek-V4.1-Flash via the DeepSeek API) on an anchored rubric, informed by deterministic scanner evidence (MCP SDK dependencies found in the repo's manifests). Prompt-injection attempts are flagged and penalized
- Score — a transparent 0-100 trust score: 35% AI rubric + 65% verifiable metrics (maintenance, popularity, docs, security posture, community). The exact formula is published in METHODOLOGY.md; every server's breakdown is in SCORES.md
- Re-evaluate — the AI re-judges servers every ~90 days; metrics and trust scores refresh every week. Projects that stagnate fall off the list
- Rank — only servers scoring 50+/100 appear here, top 20 per category, sorted by trust then stars
- Exclude — a small human-maintained exclusion list overrides the AI only for spam/scam removals and maintainer opt-outs (see CONTRIBUTING.md)
Model history: until August 2026 servers were judged by GPT-4.1-mini via GitHub Models, which GitHub shut down on July 30, 2026 (#36). From August 2026 the judge was DeepSeek-V4-Flash; since September 2026 it is DeepSeek-V4.1-Flash.
Servers are curated entirely by AI — they earn their spot through quality and lose it if they fall behind. Maintainers don't hand-pick entries. Every automated change lands as an auto-merged pull request, so the full history stays auditable and revertable.
A Note on Security
Trust scores are computed from public metadata, the README, and OpenSSF Scorecard data. Every entry's source is additionally scanned — read, never executed — for tool-poisoning markers (hidden instructions aimed at the model inside tool descriptions; flagged entries lose points in SCORES.md). Entries are keyed by GitHub's immutable repository id, so renames are followed and a known name silently re-registered by someone else (repojacking) is quarantined instead of trusted. Still: no entry has been code-audited or executed by this registry. A high score means strong public signals, not a security guarantee — review any MCP server (and the credentials you grant it) before connecting it to your tools.
Badges
Maintain a listed server? Embed your live trust score — it updates weekly and the URL survives repo renames (keyed by immutable repository id):
Your exact copy-paste snippet is under your entry in SCORES.md.
Contributing
Don't open a pull request to add a server. This README is generated from data/known_servers.json on every run, so edits to it are overwritten — and such PRs are closed automatically.
To suggest a server, open a nomination. The same AI evaluates it on the next weekly run and posts the verdict; if it scores 50+/100 it appears here automatically.
Code contributions (bug fixes, scanner improvements) are welcome — see CONTRIBUTING.md. If an AI agent is contributing on your behalf, give it AGENTS.md.
来源：GitHub · github.com