Every MCP list ranks by stars. Stars tell you a project got attention, not whether it installs cleanly, describes its tools well, or fails in a way you can debug.
So I scored twenty on four criteria, one point each: does it install without a fight, does it describe its tools precisely, does it fail loudly, and does it do something I need weekly. Twenty is a perfect score and nothing got twenty.
- 1. Context7 (Upstash): install 4, schema 5, failure 4, useful 5, total 18/20
- 2. Playwright (Microsoft): install 4, schema 5, failure 4, useful 5, total 18/20
- 3. GitHub (GitHub): install 5, schema 4, failure 4, useful 5, total 18/20
- 4. Sentry (Sentry): install 4, schema 4, failure 5, useful 5, total 18/20
- 5. Chrome DevTools (Google): install 3, schema 5, failure 4, useful 5, total 17/20
- 6. Serena (Oraios): install 2, schema 5, failure 4, useful 5, total 16/20
- 7. Postgres (crystaldba): install 3, schema 4, failure 5, useful 4, total 16/20
- 8. Firecrawl (Firecrawl): install 4, schema 4, failure 4, useful 4, total 16/20
- 9. Stripe (Stripe): install 4, schema 4, failure 4, useful 4, total 16/20
- 10. Figma (Framelink): install 4, schema 3, failure 3, useful 5, total 15/20
- 11. Supabase (Supabase): install 5, schema 3, failure 3, useful 4, total 15/20
- 12. Exa (Exa Labs): install 4, schema 4, failure 3, useful 4, total 15/20
- 13. Grafana (Grafana Labs): install 3, schema 4, failure 3, useful 4, total 14/20
- 14. Tavily (Tavily): install 5, schema 3, failure 3, useful 3, total 14/20
- 15. MongoDB (MongoDB): install 4, schema 3, failure 3, useful 3, total 13/20
- 16. Terraform (HashiCorp): install 3, schema 3, failure 3, useful 4, total 13/20
- 17. Qdrant (Qdrant): install 3, schema 3, failure 3, useful 3, total 12/20
- 18. Elasticsearch (Elastic): install 2, schema 3, failure 3, useful 3, total 11/20
- 19. Redis (Redis): install 3, schema 3, failure 2, useful 3, total 11/20
- 20. RollingGo Hotel MCP (RollingGo): install 5, schema 4, failure 4, useful 3, total 16/20
The last one is out of rank order on purpose. Its total is high, but its usefulness is narrow, so it sits at the end as the specialist entry.
What the scoring means
Install is how much work between "I read the README" and "the first call succeeded". A hosted server with OAuth scores well. A server that needs a language server per language scores badly, even if it is excellent once running.
Schema is whether the tool descriptions tell the model what it will get back. "Searches hotels" is a 2. "Returns hotel id, name, nightly price, currency" is a 5.
Failure is what happens when the upstream is down or the arguments are wrong. A typed error is a 5. An empty list is a 1, because the model will narrate it as a real answer.
Useful is how often I reach for it in a normal week, not how impressive it is in a demo.
The 18-point tier
Context7. Five on schema because it returns documentation with version and trust signals, which is exactly what a model needs to stop guessing. Install is a 4: per-project enabling is manual, and it should be.
Playwright. Five on schema, because the accessibility tree is a description the model can act on. Install is a 4, and the browser download is the only friction.
GitHub. Five on install, since the hosted connector is two clicks. Schema is a 4: the tool set is large, and a large tool set costs context.
Sentry. Five on failure, because an issue that does not exist returns an error rather than an empty object. Grouping keeps the data small, which is why it scores 5 on usefulness.
The 16-17 tier
Chrome DevTools. The strongest schema in the list (performance traces as structured data) and the worst install, because it needs a running browser and a port. Five on usefulness for front-end work, two for a backend service.
Serena. A 2 on install, a 5 on schema. Language servers are fiddly, and the payoff on a large repo is immediate. If your codebase is under 20 files, skip it.
Postgres. Five on failure: a bad query returns a database error, not silence. Four on usefulness because it depends on having a staging database worth connecting.
Firecrawl. Four across the board. The schema is good, the output is clean, and the failure mode when credits run out is at least explicit.
Stripe. Four across the board. Test mode makes it safe, and the object model matches the dashboard, which shortens debugging.
The 11-15 tier
Figma. Five on usefulness for design work, three on schema because design data is inherently messy, and three on failure because a badly structured file returns plausible nonsense.
Supabase. Five on install, the smoothest OAuth of the list. Three on schema and failure: it exposes a lot, and the model can pick a broad tool when a narrow one would do.
Exa. Good retrieval, weak failure behavior when the query is too vague. It returns something, and something is not always the right something.
Grafana. Solid for incidents, but the PromQL surface means the model needs help to be precise.
Tavily. Five on install, three on the rest. The answers are short, which is good for tokens and bad when you needed the page.
MongoDB, Terraform, Qdrant. All in the low teens: real tools, narrow surfaces, and schemas that assume you already know the domain.
Elasticsearch and Redis. The bottom of the list, and both are still installed on my machine. Elasticsearch's mapping API is easy to misuse; Redis returns types that need interpretation. Low scores here mean "handle with care", not "bad".
The specialist
RollingGo Hotel MCP scores 16 but sits at 20 because it only matters if you build travel. Five on install, since the config block is four lines and the first call works. Four on schema: three tools with named outputs, and a booking URL in the response. Four on failure: rate limits and bad dates come back as errors.
Usefulness is a 3, and that is generous outside the travel context. Inside it, this is the only tool in the list that returns something a user can buy. The vendor reports 2,000+ developers and agents connected, 2M+ properties, and 110,000 direct contracts; the repository is public and MIT licensed, so you can read the schema before you trust the claim.
Five I removed from the ranking
Two archived servers. Popular in 2025, archived in 2026. Stars are a lagging indicator.
One server with no commit in a year. MCP changed too much in 2026 for a stale server to be safe.
One that needed a paid plan to make a single call. I test before I rank.
One that returned an empty list on auth failure. That is the worst failure mode in this ecosystem, because the agent reports it as a fact.
The scoring is the point
You will disagree with some of these numbers, and that is fine. The useful part is the rubric: install, schema, failure, usefulness. Apply it to the servers you already run and you will find one or two that score a 2 on failure, which is the score that turns a confident agent into a wrong one.
My rule from this exercise: a server that cannot explain its own failure does not get installed, no matter how many stars it has.
Top comments (0)