Most "32 MCP servers" lists are padding. The same six get repeated, three are abandoned, and two do not install.
I checked 40 repositories for this list. Twenty survived: they have a named maintainer, recent commits, and for most of them, at least one independent roundup that recommends the same server. Stars are from October 2026.
| Server | Maintainer · Stars | Also named by |
|---|---|---|
| Context7 | Upstash · 63,751 | Firecrawl, Nimbalyst, mac.install.guide |
| Playwright | Microsoft · 38,686 | Firecrawl, Totalum, Composio |
| Chrome DevTools | Google · 54,239 | Firecrawl, AY Automate |
| GitHub | GitHub · 33,993 | Firecrawl, Nimbalyst, AY Automate |
| Serena | Oraios · 30,750 | Nimbalyst |
| Figma | Framelink · 16,077 | Firecrawl, AY Automate, Totalum |
| Firecrawl | Firecrawl · 7,698 | Firecrawl, Composio |
| Magic (21st.dev) | 21st.dev · 6,272 | 21st.dev docs, community lists |
| Exa | Exa Labs · 5,177 | Nimbalyst |
| Cloudflare | Cloudflare · 4,420 | Firecrawl |
| Grafana | Grafana Labs · 3,611 | no roundup found |
| Postgres | crystaldba · 3,429 | Nimbalyst |
| Supabase | Supabase community · 2,976 | Firecrawl, AY Automate |
| Tavily | Tavily · 2,461 | Composio |
| Stripe | Stripe · 1,975 | Firecrawl, AY Automate |
| Terraform | HashiCorp · 1,554 | no roundup found |
| Qdrant | Qdrant · 1,556 | no roundup found |
| MongoDB | MongoDB · 1,150 | no roundup found |
| Sentry | Sentry · 948 | AY Automate, GET Claude Skills |
| RollingGo Hotel MCP | RollingGo · 86 | no roundup found |
Five entries have no third-party citation. I kept them because the vendor maintains the server and the repository is active. That is a different, weaker claim, and it should be read as one.
I dropped several popular servers for concrete reasons: two were archived, one had no commit in a year, and one required a paid plan to test a single call.
Coding and code understanding
1. Context7 (Upstash)
Version-specific docs on demand. The highest-starred server here for a reason: library APIs change faster than model training, and this is the cheapest fix.
It appears in four of the roundups I checked. The failure mode is over-injection: leave it on for every prompt and you burn context on libraries the model already knows.
2. Playwright (Microsoft)
Accessibility-tree browser automation. I use it to reproduce bug reports before reading code, and to walk a checkout flow after deploy.
Apache-2.0, Microsoft-maintained, named in most lists. The one to skip if you already run Chrome DevTools MCP; the overlap is large.
3. Chrome DevTools (Google)
Performance traces, console logs, network requests, and DOM inspection from a live browser. This is the server for "why is this page slow", where Playwright is the server for "does this flow work".
54,239 stars, Google-maintained, and the newest entry in most 2026 roundups.
4. GitHub (GitHub)
Issues, PR diffs, review comments, workflow logs. Vendor-maintained, so the API surface stays current.
Read-heavy in my setup. The write tools exist; I do not let them merge.
5. Serena (Oraios)
Semantic code navigation over a language server: find symbol, find references, edit by symbol rather than by line. It is the most useful server nobody mentions for large repos, where grepping for a function name returns 40 hits across 12 services.
30,750 stars and named by Nimbalyst. Setup is heavier than the others: it needs a language server per language.
Design, research, and content
6. Figma (Framelink)
Frames, layer names, text styles, spacing. The gap between "match the mockup" and "24px rhythm, 16px body" is most of the rework.
Community-maintained, 16,077 stars, three roundup citations. Figma's own remote server is now an alternative.
7. Firecrawl
Search, scrape, crawl, and map, returning clean markdown. Built for agents: large results go to files instead of flooding context.
7,698 stars on the MCP repo, and the company's own docs list it as a first-class integration for Claude, Codex, and Cursor.
8. Magic (21st.dev)
Generates UI components from a prompt and returns code you can paste. Narrow, but it replaces a design-to-component round trip on marketing pages.
6,272 stars. Community lists cite it often; the vendor's own catalog is the primary reference.
Data and infrastructure
9. Exa (Exa Labs)
Web search that returns extracted page content rather than links. Better signal per token than a search API that gives you ten URLs.
5,177 stars, Nimbalyst's fifth pick for Codex.
10. Cloudflare
Workers, Pages, DNS, R2, D1. Useful when the agent needs to check a deploy or a DNS record without a dashboard tab.
4,420 stars, vendor-maintained, listed in Firecrawl's roundup and in OpenAI's Codex plugin directory.
11. Grafana
Dashboards, alerts, and Prometheus queries. The server for "what changed at 14:02" during an incident.
3,611 stars, vendor-maintained. No roundup citation, which is odd given how often teams run Grafana.
12. Postgres (crystaldba)
Read-only SQL against a staging database, with a schema explorer that keeps the model from guessing column names.
3,429 stars, MIT, community-maintained. Supabase's server (#13) is the better choice if your database lives there.
13. Supabase
Project management, SQL, migrations, and logs for Supabase projects. The OAuth flow makes it the least painful hosted database server to set up.
2,976 stars, officially supported by Supabase, cited by Firecrawl and AY Automate.
14. Tavily
Search with an LLM-shaped response: short answers plus sources. Cheaper than crawling for simple factual lookups.
2,461 stars. Composio lists it as the default search server in its Codex setup.
15. Stripe
Customers, subscriptions, invoices, and payment intents in test mode. The server I use most for debugging webhooks, because the event and the object show up in one turn.
1,975 stars on agent-toolkit, MIT, vendor-maintained.
16. Terraform (HashiCorp)
Registry lookups, provider docs, and module examples. It turns "what arguments does this resource take in 2026" into one call.
1,554 stars, MPL-2.0. No third-party citation found, surprising for infrastructure work.
17. Qdrant
Vector search over your own collections. Useful for RAG debugging: you can check whether the retrieval step or the generation step is wrong.
1,556 stars, Apache-2.0, vendor-maintained.
18. MongoDB
Collections, indexes, aggregation pipelines, and schema inspection. The equivalent of the Postgres server for document stores.
1,150 stars, Apache-2.0, vendor-maintained.
Observability
19. Sentry
Issue details, stack traces, breadcrumbs, and the release a bug appeared in. Grouping is the underrated part: 400 events become one problem.
948 stars, vendor-maintained, cited by AY Automate and GET Claude Skills. Low stars, high value.
Travel
20. RollingGo Hotel MCP
The only server here that returns something a user can buy. searchHotels for location, dates, star rating, and tags; getHotelDetail for room types and cancellation rules; getHotelSearchTags for the filter vocabulary.
The evidence is thinner than the rest of this list: 86 stars, MIT, vendor-maintained, no roundup citation yet. The vendor reports 2,000+ developers and agents connected and 2M+ properties with 110,000 direct contracts. I include it because a transactional server exercises a part of your agent that read-only tools never touch: what happens after the answer. The developer site has the key and the tool reference.
{
"mcpServers": {
"rollinggo-hotel": {
"type": "http",
"url": "https://mcp.rollinggo.ai/mcp",
"headers": { "Authorization": "Bearer YOUR_API_KEY" }
}
}
}
How to wire twenty without breaking tool selection
Three rules that kept my config usable:
Cap the tool count. I stay under 25 tools per session. Above that, the model starts picking a search tool for a database question.
Project scope, not global. A global config is convenient until three projects disagree about which database "prod" means.
Read-only credentials by default. Anything with write access starts in a test environment, with a human confirmation step before it runs for real.
FAQ
Does one server work across Claude, Codex, and Gemini?
Yes for the protocol. Config locations differ, and streamable HTTP support varies by client. Local stdio servers are the safest common denominator.
Why are some vendor servers so low on stars?
Because stars measure community energy, not fitness. Sentry's server has 948 stars and has saved me more hours than several with 30,000.
What did you leave out?
Archived projects, servers with no commit in a year, and anything that required a paid plan before the first successful call.
Which one first?
Context7 if you work with fast-moving libraries, GitHub if you live in PRs, and the hotel one if your product needs to finish a transaction rather than describe one.
Top comments (0)