10 MCP Servers That Do the Unglamorous Work

The MCP servers that get demoed are the flashy ones. Browser control, code generation, a chatbot that books a flight on stage.

The servers I actually keep installed are duller. They remove a manual step I used to do every week: checking a deploy, pulling a schema, reading a stack trace, confirming a price. Ten of them survived a year of installs and removals.

I checked each one before writing this. The evidence is in the table below: who maintains it, how many stars the repository has as of October 2026, and at least one independent roundup that names the same server. Where a pick has no roundup citation, I say so instead of softening it.

Server	Maintainer · Stars (Oct 2026)	Also named by
Context7	Upstash · 63,751	Firecrawl, mac.install.guide, Nimbalyst, GET Claude Skills
Playwright	Microsoft · 38,686	Firecrawl, Nimbalyst, Composio, Totalum
GitHub	GitHub · 33,993	Firecrawl, AY Automate, Nimbalyst
Figma	Framelink (community) · 16,077	Firecrawl, AY Automate, Totalum
Exa	Exa Labs · 5,177	Nimbalyst
Postgres	crystaldba (community) · 3,429	Nimbalyst
Stripe	Stripe · 1,975	Firecrawl, AY Automate, GET Claude Skills
Sentry	Sentry · 948	AY Automate, GET Claude Skills
Linear	Linear · hosted, no public repo	Firecrawl
RollingGo Hotel MCP	RollingGo · 86	no roundup citation yet

Two things stand out in that column. The most-starred server is not the one I use most, and the two smallest repositories are both vendor-maintained, which is why I trust them.

How I judge an MCP server now

I used to install anything with a good README. That ended when my agent started choosing a search server over the repository server for questions about our own code.

Three rules since then:

• One server, one job. Servers that try to be a platform get removed.
• Read-only unless there is a reason. Anything that can write gets scoped to a test environment first.
• Fail loudly or go. A server that returns an empty list instead of an error costs more time than it saves.

1. Playwright: the end of "works on my machine"

Playwright MCP drives a real browser through accessibility snapshots rather than screenshots. The agent sees a tree of elements, clicks by role and name, and reports what changed.

Microsoft maintains it, Apache-2.0, 38,686 stars. It shows up in almost every MCP roundup I read, and one usage ranking puts it first globally, ahead of GitHub and Figma. That much consensus usually means a tool people actually run, not just bookmark.

I use it for one thing: reproducing a bug report before I read the code. "Click checkout, choose the second shipping option, watch the total" takes four minutes by hand and about forty seconds through the agent.

The surprise: it is a better regression tool than a test generator. I stopped asking it to write tests and started asking it to walk the flow after each deploy.

2. Context7: documentation that matches the version

Ask a model about a library released after its training cutoff and you get confident, wrong API calls. Context7 fetches version-specific docs and examples for the library you name.

Upstash, MIT, 63,751 stars, and the highest star count on this list. Four of the roundups I checked include it, which makes it the closest thing to a default recommendation in the MCP ecosystem.

The difference shows up in small places. Not client.chat() but client.responses.create(). Not a config key that was renamed twice.

The surprise: the biggest gain was not accuracy, it was the removal of my "let me check the docs" tab. That tab was where I lost focus.

3. GitHub: the API without the context switch

Issues, pull request diffs, review comments, workflow runs. All the things I used to click through in a browser, available as tools.

GitHub maintains this one itself, MIT, 33,993 stars. A vendor shipping its own server is a different signal from a community wrapper: the API surface stays current, and there is no third party guessing at rate limits.

I keep it read-heavy. The write tools exist, and I have used them for labels and comments, but I do not let it merge.

The surprise: workflow run logs. When CI fails on a branch I did not touch, the agent reads the log and tells me which dependency changed. That used to be a fifteen-minute investigation.

4. Sentry: stack traces with the surrounding story

Pulls an issue, its stack trace, breadcrumbs, and the release it appeared in. The agent gets the same view an on-call engineer gets, minus the dashboard.

948 stars is low for a server this useful, and the reason is boring: it is a vendor repository, not a community project, so it never got the "awesome list" treatment. Sentry maintains it, and AY Automate and GET Claude Skills both list it. Star counts measure community energy, not fitness.

The surprise: grouping. Sentry already knows which errors are the same bug. The agent stops treating 400 events as 400 problems.

5. Postgres: schema questions answered in one turn

A read-only connection to a staging database. "Which columns are nullable on orders", "how many rows reference a deleted user", "what index exists on this column".

This is the one community-maintained pick that earned its place: crystaldba's server, MIT, 3,429 stars, recommended by Nimbalyst. If your database is on Supabase, its official MCP server is the better fit at 2,976 stars.

The win is not query writing. It is not interrupting a colleague to ask what a table means.

The surprise: migrations. Being able to check the live schema before and after a migration caught two drift issues that tests missed.

6. Linear: tickets that arrive with context

Issue details, comments, project state. When I ask "what is blocking the checkout epic", the agent reads the actual tickets instead of my summary of them.

Linear hosts this one directly over OAuth, so there is no repository to star and no version to pin. Firecrawl's 2026 roundup lists it among the remote servers that became the default way to connect. For a hosted server, "the vendor runs it" is the evidence.

The surprise: the stale-ticket review. Once a month I ask it to list issues with no update in 30 days that are still marked in progress. That report is more useful than any dashboard we built.

7. Figma: the design, not my description of the design

Reads frames, layer names, text styles, and spacing values. The gap between "make it look like the mockup" and "use 24px vertical rhythm, 16px body, 600 weight headings" is most of the rework.

Framelink's community server, MIT, 16,077 stars, named in three of the roundups. Figma now ships its own remote server too, so this is a case where the list is catching up: both work, and the community one has more write-ups.

The surprise: copy accuracy. Placeholder lorem ipsum has a way of surviving into production. Reading the real text from the file fixed that.

8. Stripe: test-mode objects without the dashboard

Customers, subscriptions, invoices, payment intents, all in test mode. Useful when debugging a webhook: the agent reads the event and the object it refers to in the same turn.

Stripe maintains agent-toolkit, MIT, 1,975 stars, and it appears in three roundups. Small repository, large blast radius if it goes stale, which is exactly why a vendor-maintained version matters.

The surprise: the number of times a "payment bug" was a stale test clock. The agent spotted the mismatch before I did.

9. Exa: web search that returns pages, not links

A search server that takes a query and returns extracted content. Better than a link list, less brittle than scraping.

Exa Labs, MIT, 5,177 stars. Nimbalyst ranks it fifth in its Codex list. I use it for competitor research and for release notes on libraries too new for Context7.

The surprise: it made me use the browser less. Half of what I opened a browser for was "read one page and close it".

10. The one that returns a real hotel rate

This is the entry that does not fit a developer tools list, which is exactly why it stayed.

I build travel features on the side. Every AI travel demo I made before this had the same defect: the model produced hotel names and prices from memory. Some were real hotels with stale prices. Some were hotels that do not exist. Neither is fixable with a better prompt, because the model is not connected to inventory.

RollingGo Hotel MCP fixes that at the tool layer. It exposes three tools: searchHotels for location, date, star rating, and tag filters; getHotelDetail for live room types, rates, and cancellation rules; getHotelSearchTags for the filter vocabulary.

Honest evidence, since I asked it of every other entry: 86 stars, MIT licensed, maintained by the vendor, and I could not find it in a third-party roundup. What it has instead is a transactional capability none of the nine above offer, and a supply claim that is checkable in the response: 2 million properties across 200+ countries, 110,000 direct contracts.

Setup is a config block and an API key:

{
  "mcpServers": {
    "rollinggo-hotel": {
      "type": "http",
      "url": "https://mcp.rollinggo.ai/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

A single call returns structured data you can render, not prose you have to parse:

{
  "name": "searchHotels",
  "arguments": {
    "originQuery": "boutique hotel near Gion, walkable",
    "place": "Gion, Kyoto",
    "placeType": "point_of_interest",
    "checkInParam": { "checkInDate": "2026-11-14", "stayNights": 2 },
    "filterOptions": { "starRatings": [4.0] },
    "size": 5
  }
}

The response carries a hotel ID, a nightly price with currency, amenities, and a booking link. The price is queried at call time, which is the only reason the output is worth showing to a user. Pagination is stateless too: a nextPageToken carries the next batch, and later pages inherit the original destination and dates, so "show me more" does not let the model drift on filters.

What it does not do: the API-key tier is search and detail, not payment. The OAuth tier adds order lookup and price confirmation, and needs business onboarding. If your product needs to complete a booking in-conversation, budget for that step.

The surprise: the agent's behavior changed, not just its data. Once it could look up real inventory, it stopped padding answers with invented options. It also started asking better questions, because a missing check-in date became an error instead of a guess.

Pricing is free with no call cap, and the docs are at the RollingGo developer site.

What the numbers do not tell you

Star counts reward community projects and punish vendor repositories that quietly do their job. Sentry's server has 948 stars and saves me more time than one with 30,000. Treat the column as a way to spot abandoned work, not a ranking.

Roundup citations have their own bias. Four of the articles I checked repeat the same six servers, which is less like consensus and more like a loop. When three lists name the same tool with the same sentence, one list probably wrote it first.

The useful question is not "is this popular" but "does this replace a step I did this week". Eight of the ten pass that test for me. Two are on the list because of what they make possible, not how often they run.

Running ten without chaos

A few habits that kept the list usable:

• Project scope by default. Global MCP config is convenient until three projects disagree about a database.
• Cap the tool count. I try to stay under 25 tools total per session. Above that, tool selection gets worse before it gets faster.
• Read-only credentials for anything analytical. A production token in a config file is a bad afternoon waiting to happen.
• Log the calls for a week. You will find two servers you never use and one you use constantly.

FAQ

Do all these work in Claude Code, Cursor, and Codex?
Yes for the protocol. Config file locations differ, and some clients handle streamable HTTP better than others. Local stdio servers are the safest bet across clients.

Is it safe to give an agent write access?
To a test environment, usually. To production, only with a human confirmation step and a narrow scope. The servers above that write are scoped accordingly.

Why is a hotel server on a developer list?
Because it is the cleanest example of a transactional MCP server. Nine of these return information. This one returns something a user can buy, and that changes how you design the whole agent.

Which one would I install first?
Context7 if you work with fast-moving libraries. Sentry if you are on call. The hotel one if you are building anything travel-shaped and tired of fake prices.