MCP server categories

These servers let an AI assistant query and manage real databases: run SQL, inspect schemas, and answer questions from live data instead of guesses. Most support read-only modes, which is the right default for anything an agent touches.

Language models stop knowing things the day their training ends. These servers wire an assistant to live search engines and web content, so it can look things up, cite sources, and stop hallucinating yesterday's facts.

These servers give an agent hands on the web: load pages, click, fill forms, take screenshots, and extract data from sites that have no API. If the task starts with 'go to this website and…', this is the category.

The largest category in the catalog, because this is where agents earn their keep: repos, pull requests, pipelines, containers, and cloud resources. Watch the permissions panel closely here; write access to infrastructure deserves scrutiny.

Incident response is a search problem, and agents are fast searchers. These servers connect assistants to logs, metrics, error trackers, and traces, so 'why is prod slow' becomes a query rather than a scavenger hunt.

The bread and butter: reading, writing, and organizing files, locally or in cloud storage. Sandboxing matters most here; the best servers scope an agent to the directories you choose and nothing else.

Servers that read payment processors, invoices, and accounting systems, so an assistant can answer revenue questions and chase paperwork. Read-scoped access is the norm in the good ones, and that's how you should run them.

On-chain data is public but hostile to read. These servers let an agent query blockchains, wallets, and protocols directly, turning addresses and transactions into answers. Treat anything that can sign or send with extreme care.

Where your team already talks: Slack, email, and messaging platforms. These servers let an assistant search history, summarize threads, and draft replies. The permissions panel tells you whether it can only read or also speak as you.

Agents calling other models: embeddings, image generation, transcription, and specialized inference. Useful when one assistant orchestrates and the heavy lifting happens in purpose-built models behind these servers.

Calendars, task boards, notes, and docs: the glue of the workday. These servers let an assistant read what's planned and update what's done, which turns 'remind me' and 'schedule it' from wishes into tool calls.

Servers that process and produce media: image manipulation, video, audio, and design tool integrations. The practical wins are pipeline chores, such as conversions, crops, and metadata, done by an agent instead of a human with twelve tabs open.

Scanners, advisory feeds, and analysis tools an agent can drive. Security work is checklist-heavy, and agents are relentless about checklists. Ironically, this is also the category where you should read the permissions panel twice.