The first hundred¶
Companion to skill-catalogue.md, which decided how the catalogue works.
This decides what goes in it, and in what order.
"Most used" is not published — this is triangulated¶
No one publishes MCP server usage. Every ranking online is a proxy, and the proxies disagree because they measure different things. Six sources, and what each actually measures:
| source | measures | caveat |
|---|---|---|
| PulseMCP — 22,000+ servers, aggregates the official registry, Glama, Smithery, mcp.so | visitors to the server's listing | interest, not installs |
| mcpmanager.ai top 50 | search volume (622k/mo worldwide) | demand, and inflated for servers whose names are ordinary words |
| best-of-mcp-servers | GitHub stars | often the parent project's stars — microsoft/markitdown at 180k is a document converter that also ships an MCP server |
| modelcontextprotocol/servers | official reference set | authoritative, and explicitly "educational examples", not production |
| mcpservers.org remote list | vendor-hosted endpoints | the best signal for maintained, which matters more than popular |
| Docker MCP catalog — 300+ verified | passed a review, signed, SBOM | quality, not usage |
Where they agree, the signal is real: Playwright, GitHub, Filesystem, Fetch, Context7, Notion, Supabase, Linear, Slack, Figma and Sequential Thinking appear near the top of every list. Where they disagree, this document says so rather than averaging them into false precision.
One pattern worth naming: the top of every ranking is browser automation, documentation lookup and code tools. That is the coding-agent workload, because coding agents are who adopted MCP first — not evidence that nobody wants a CRM.
The finding that shapes the catalogue¶
Sampling 76 servers that recur across these sources, and asking the only question our liveness check cares about — can we start this without somebody's account?
21 28% verifiable as-is starts with no secret
10 13% verifiable in CI with a container we can stand the dependency up ourselves
45 59% unverifiable in public CI needs someone's account
Roughly three in five of the most-wanted servers can never carry a verified badge in public CI.
That is not a problem to design around; it is the thing to be honest about, and it is why
unverifiable is a first-class state rather than a silent pass.
It also produces one concrete recommendation. The 13% — Postgres, Redis, MongoDB, Elasticsearch,
Neo4j, Qdrant, ClickHouse, Kubernetes — are unverifiable only because nothing stands the dependency
up. GitHub Actions services: can. Containerising those moves a seventh of the catalogue from
"we did not look" to "verified on a date", and they are exactly the servers where a silent schema
change is most likely to hurt.
Tier A — verify as-is (seed the catalogue with these)¶
No credential, no container. The liveness check works on day one, so the badge means something before anything else is added.
| # | server | maintainer | why it is here |
|---|---|---|---|
| 1 | Playwright | Microsoft | #1 or #2 in every source; 76.9m PulseMCP visitors, 82k monthly searches |
| 2 | Chrome DevTools | 51.2m visitors — second only to Playwright, and absent from the search-volume list, which is why one source is not enough | |
| 3 | Fetch | Anthropic | official reference; 32.2m visitors |
| 4 | Filesystem | Anthropic | official reference; 16m visitors |
| 5 | Git | Anthropic | official reference; 9.1m visitors |
| 6 | Sequential Thinking | MCP steering group | official; most-installed on Smithery (5,550+) |
| 7 | Memory / Knowledge Graph | Anthropic | official reference; 4.1m visitors |
| 8 | Time | Anthropic | official reference; 6.1m visitors |
| 9 | Context7 | Upstash | 33.5m visitors; live library docs — the highest-value non-Anthropic free entry |
| 10 | markitdown | Microsoft | document → markdown; already used in this repo |
| 11 | DuckDB | ktanaka101 | 4.9m visitors; embedded, so genuinely no dependency |
| 12 | SQLite | community | same argument as DuckDB |
| 13 | Puppeteer | community | 7.3k searches; the alternative when Playwright is too heavy |
| 14 | Browser Use | Browser Use | 31.1m visitors |
| 15 | Serena | oraios | 19k searches, 29k stars; symbolic code operations |
| 16 | Everything / Demo | Anthropic | the reference test server — the natural fixture for our own check |
| 17 | Office Word | gongrzhe | 5.7m visitors |
| 18 | Office PowerPoint | gongrzhe | 1.9m visitors |
| 19 | XcodeBuild | Cameron Cooke | 2.4m visitors |
| 20 | Storybook | Storybook | 29.8m visitors |
| 21 | Blender | Siddharth Ahuja | 2.8m visitors — high interest, narrow audience |
Tier B — verifiable with a container in CI¶
Unverifiable only because nothing starts the dependency. services: in the workflow fixes that, and
these are where an unnoticed schema change does the most damage.
| # | server | note |
|---|---|---|
| 22 | PostgreSQL | 7.9k searches, 3.7m visitors; the most-wanted database |
| 23 | MySQL | 4.2k searches |
| 24 | Redis | 1.4k searches |
| 25 | MongoDB | official (MongoDB Inc.), 3.2m visitors |
| 26 | Elasticsearch | Docker-verified partner |
| 27 | ClickHouse | analytics workloads |
| 28 | Neo4j | graph queries |
| 29 | Qdrant | vector search |
| 30 | Chroma | vector search, embedded option |
| 31 | Kubernetes | 2.1k searches |
| 32 | Docker | 10.3k searches |
Tier C — vendor-hosted and vendor-maintained (unverifiable, low rot)¶
Credentialed, so permanently unverifiable in public CI — but the vendor maintains them, which is
the risk that actually matters. A vendor-run endpoint is far less likely to silently rename a tool
than a community wrapper around the same API.
Development: GitHub · GitLab · Linear · Atlassian (Jira/Confluence) · Sentry · Vercel · Railway · Neon · Supabase · Cloudflare · AWS · Azure · Docker Hub · CircleCI · Netlify
Product and design: Figma · Canva · Notion · Storybook · PostHog · Amplitude
Communication and work: Slack · Asana · Monday · ClickUp · Trello · Box · Google Drive · Gmail · Google Calendar · Google Sheets · Microsoft 365 · Outlook
Commerce and finance: Stripe · Shopify · Mercury · Plaid · QuickBooks · Xero · PayPal
CRM and marketing: Salesforce · HubSpot · Attio · Apollo.io · Intercom · Zendesk · Mailchimp
Data and observability: Snowflake · Databricks · dbt · Datadog · Grafana · New Relic · Elastic Cloud · BigQuery
Automation: Zapier · n8n · Activepieces · Make
Search and retrieval: Brave Search · Tavily · Exa · Firecrawl · Perplexity · Kagi · SerpAPI · Parallel Search · Ahrefs
That is roughly the remaining 70, and the count is deliberately approximate: the exact tail matters less than the ordering, and the tail should be driven by what people ask for rather than by what a directory happened to list.
What to actually build first¶
Do not seed with a hundred. Fifteen that provably work beats a hundred that might, and the whole argument for building verification before the repo was that the badge should mean something on day one.
Seed with Tier A 1–16 — the sixteen that need no credential and no container. That is:
- every official Anthropic reference server (7)
- the two browser servers at the top of every ranking
- Context7, the highest-value free third-party entry
- markitdown, DuckDB, SQLite, Puppeteer, Serena, Browser Use
All sixteen verify on day one. The catalogue launches green, honestly.
Then add Tier B, once the workflow grows services: — and that PR is worth doing for its own
sake, because it converts a seventh of the catalogue from a shrug into a fact.
Then Tier C, driven by requests. The ordering within it is a guess; the first issues will be better evidence than any directory.
Open questions this raises¶
- Remote/hosted servers do not fit the current shape. Tier C is mostly OAuth endpoints
(
https://mcp.linear.app/mcp), not stdio commands.mcp_serversupportstransport: http, but the credential story for a catalogue entry — how a user supplies their own token at import — is unresolved. - A bundle's floor and state. Per the "bundle wraps skills" decision: effective floor is the max of its members', verification state the worst. Neither is implemented, because bundles do not exist yet.
- What is a skill here, versus what is a tool? GitHub's server exposes 51 tools. A bundle with 51 skills is not useful to an agent; picking 6 is a curation judgement this document does not make.
- Licence and attribution. Carrying a third party's config is not carrying their code, but the catalogue should record the upstream licence per entry anyway.