"Hidden" is doing a lot of work in most lists like this. They mean "not the six everyone repeats", and then they recommend a server with 40,000 stars that appeared in every roundup last month.
My rule for this list: under 6,000 stars, actively maintained, and doing a job that comes up in a normal week. That filters out both the overexposed and the abandoned. Two servers I wanted to include were archived, one had no commit since 2025, and those are the reasons they are not here.
| Server | Job · Maintainer | Stars · Roundup citation |
|---|---|---|
| Postgres | Query a staging DB · crystaldba | 3,429 · Nimbalyst |
| Grafana | Dashboards and alerts · Grafana Labs | 3,611 · none found |
| Tavily | Search answers with sources · Tavily | 2,461 · Composio |
| Terraform | Provider and module docs · HashiCorp | 1,554 · none found |
| Qdrant | Vector search debugging · Qdrant | 1,556 · none found |
| DuckDuckGo | Free web search · community | 1,553 · none found |
| Exa | Semantic web research · Exa Labs | 5,177 · Nimbalyst |
| MongoDB | Document store queries · MongoDB | 1,150 · none found |
| Sentry | Runtime error context · Sentry | 948 · AY Automate |
| Elasticsearch | Log and index search · Elastic | 725 · none found |
| Redis | Cache and key inspection · Redis | 641 · none found |
| MotherDuck | DuckDB in the cloud · MotherDuck | 535 · none found |
Six of the twelve have no third-party citation. Vendor-maintained and recently shipped is the claim I can defend, so that is the claim I am making.
The four I reach for weekly
1. Postgres (crystaldba)
Read-only SQL against staging with a schema explorer. The questions it answers are the ones that cause silent bugs: which columns are nullable, which index exists, whether a foreign key actually points where the diagram says.
3,429 stars, MIT, community-maintained, cited by Nimbalyst. If your database is Supabase, use their official server instead; it is more integrated and equally safe when scoped correctly.
2. Grafana
Dashboards, alert rules, and PromQL queries. During an incident, the agent can ask "what changed at 14:02" and read the same dashboard you would open.
3,611 stars, vendor-maintained. No roundup citation, which surprises me given how many teams run Grafana. This is the kind of tool that gets used and never written about.
3. Sentry
Issue details, stack traces, breadcrumbs, and release info. 948 stars and one of the highest value-to-star ratios on the list.
Vendor-maintained, cited by AY Automate and GET Claude Skills. The grouping is the underrated part: the model stops treating 400 events as 400 problems.
4. Tavily
Search with short answers and sources, designed for agents rather than humans. Cheaper than crawling when you need a fact, worse than Firecrawl when you need a page.
2,461 stars. Composio lists it as a default search server in its Codex setup.
Infrastructure and data
5. Terraform (HashiCorp)
Registry lookups, provider docs, and module examples. "What arguments does this resource take in the current provider version" is a question models get wrong constantly, because provider schemas change every few weeks.
1,554 stars, MPL-2.0, vendor-maintained.
6. Qdrant
Vector search over your own collections. I use it for RAG debugging: given a bad answer, check whether retrieval returned the right chunk before blaming the model.
1,556 stars, Apache-2.0, vendor-maintained.
7. MongoDB
Collections, indexes, aggregation pipelines, and schema sampling. Document stores punish guessing more than relational ones, because there is no schema to read from the code.
1,150 stars, Apache-2.0, vendor-maintained.
8. Elasticsearch
Index mapping, query DSL, and log search. The mapping question alone, "is this field analyzed or keyword", explains a class of search bugs that look like application logic.
725 stars, Apache-2.0, vendor-maintained.
9. Redis
Key inspection, TTLs, and memory usage by prefix. Useful when the bug is "the cache is stale" and nobody wants to open a CLI at 11 p.m.
641 stars, MIT, vendor-maintained.
10. MotherDuck
DuckDB in the cloud: run analytical SQL against a warehouse without exporting a CSV first. 535 stars, MIT.
This is the smallest project on the list and the one I use least, but when the question is "what do the numbers actually say", it saves a round trip through a BI tool.
Research
11. Exa
Semantic search that returns extracted page content. Better than keyword search for "find the changelog entry that renamed this option".
5,177 stars, MIT, Nimbalyst's fifth pick for Codex. The failure mode is over-fetching; two results, not ten.
12. DuckDuckGo
Free, no API key, no account. When you need one page and do not want to wire a paid provider, this is the server that gets you there in a minute.
1,553 stars, community-maintained. Quality is below Tavily and Exa; the price is zero and so is the setup.
What I left out, and why
Archived repositories. Two popular servers I wanted to include are archived. An archived server still works today and will not work after the next API change.
Anything that needed a paid plan to run one call. I test before I recommend. If the first call requires procurement, it is not a hidden gem, it is a sales funnel.
Servers with no commit in a year. MCP moved a lot in 2026. Stale is a synonym for broken in a few months.
The big six. Context7, Playwright, GitHub, Figma, Firecrawl, and Chrome DevTools are excellent and appear in every other list. Excluding them is the point of this one.
How to add one without wrecking tool selection
Three rules from a year of trial and error:
One server, one job. Two search servers means the model picks one at random and you cannot tell which.
Project scope. A global config is convenient until three repos disagree about which database is "prod".
Read-only credentials. Every server above that touches data starts with a read-only role. The one time I skipped that, the agent helpfully "fixed" a staging table during a demo.
Which one first
If you are on call: Sentry, then Grafana, then Postgres. That trio answers "what broke, when did it change, and what does the data say" without a single dashboard tab.
If you are doing data work: Postgres, then Qdrant, then MotherDuck.
If you are building a product with a transactional step, none of these will help, and that is a different category of server entirely. A read-only tool makes your agent accurate. A transactional one makes it useful, and it needs a much stricter review of what it is allowed to do. The one I use for that is the RollingGo Hotel MCP server, which returns real rates and a booking link instead of a description.
Top comments (0)