Our MCP catalogue

MCP servers for Claude Code and Claude Cowork that we run daily on client work, plus MCP tooling for building and inspecting coding agents.

An MCP (Model Context Protocol) is a server that gives Claude access to an outside tool, in Claude Code and in Claude Cowork alike: reading a Trello board, running an Apify scraper, checking a project's secrets in Doppler. A Claude Code skill is a different thing: a folder of instructions and scripts, with no external server involved; that is the difference between skills and MCP in Claude Code. We run both together on most projects: the MCP connects the tool, and the skill decides when and how to use it.

Official MCPs

06

Official third-party MCP servers we connect to Claude Code and Claude Cowork for day-to-day client work.

Apify

Apify's official MCP for running Actors (scrapers and crawlers) from Claude.

Useful forWe use it as the scraping layer of our SEO analysis skill, for keyword volumes and SERPs when Google's official API doesn't cover the data. We pick it over building our own scraper because the Actor already exists, someone else maintains it, and Claude can call it directly as just another MCP tool.

Connects Claude Code or Claude Desktop to the Apify Store, so an agent can find an Actor, run it with the parameters it needs and read back the dataset, without leaving the conversation. It's a remote HTTP server with OAuth against your Apify account, or a plain API token if you'd rather skip the browser. Every run spends credits from your Apify plan, same as launching it from their console.

Descript

Descript's official MCP for editing video and audio from the assistant.

Useful forWe use it to clean up consulting-session recordings and product clips before publishing, asking Claude for the edits instead of opening the editor by hand. Only a file that's already published or exported is visible to the agent, everything else still gets edited inside Descript directly.

Gives Claude access to one Descript Drive (you pick which during auth) to import media from a URL or a local file, which gets transcribed automatically. From there it can ask Underlord to strip filler words, apply Studio Sound, add captions or suggest B-roll, and create or manage projects by prompt. It's a Descript-hosted server, no API key or local process to keep running.

Doppler

Doppler's official MCP for reading projects, configs and secrets from the agent.

Useful forWe use it on projects where the agent needs real environment variables (an API key, a connection string) without that secret ever sitting in a file in the repo. We install the server read-only by default, dropping that flag is the exception and a deliberate call.

Doppler keeps secrets off disk and injects them as environment variables when a process starts, and this MCP lets Claude query that structure (projects, configs) instead of a local `.env`. Setup means creating an isolated config for the agent, granting only the secrets the task needs, and generating a read-only, expiring service token. The project and environment the MCP uses live in the folder binding `doppler setup` creates, not in its own config file.

Fathom

Fathom's official MCP for reading meeting notes and transcripts.

Useful forWe use it right after a client call so Claude can draft the minutes or pull out next steps straight from the transcript, without anyone re-listening to the recording. We pick it over asking Claude to summarize from memory because every claim traces back to the actual transcript, not a guess.

Connects Claude to your Fathom account so it can list meetings, and for a specific call pull its summary or the full timestamped transcript. Authorization is browser OAuth, and on team accounts an owner has to enable the connector before it shows up for everyone else.

Tavily

Tavily's official MCP for web search, extraction, mapping and crawling.

Useful forWe use it in our research skills when the task needs to crawl a whole domain or pull the full text of several pages, not just fire off one search. We pick it over Claude's built-in `WebSearch` when we need that structured mapping or extraction; for a single lookup, the built-in tool is enough.

Gives Claude four web tools beyond a plain search: search, extract the clean content of a URL, map a site's structure, and crawl multiple pages in depth. It's set up with a free Tavily API key, either as a local server via npx or as a direct URL connection with the key baked into the address.

Trello

Trello's (Atlassian) official MCP for reading and writing boards, lists and cards.

Useful forWe use it so Claude turns meeting notes or an email into cards already filed on the right board, instead of someone copying them over by hand afterward. On projects that already run a different task manager, we skip it, the gain only shows up when Trello is the team's real board.

Connects Claude directly to your Trello workspace: it can read boards, lists, cards and checklists, and create or move cards on request. It's a remote server with OAuth, no local code or API key to manage.

MCP for coding agents

02

Model Context Protocol servers and SDKs built for constructing and inspecting coding agents, not for end use.

MCP-UI

An SDK for rendering interactive UIs inside MCP tools.

Useful forBuilding your own MCP servers whose tools return forms, dashboards or widgets rather than text, rendered inside hosts like Claude. It's an MCP-building block, not an MCP we install for daily use.

Lets developers attach interactive web interfaces to MCP tools, so AI hosts can render rich UIs alongside tool output instead of plain text. It ships server SDKs (TypeScript, Ruby, Python) for creating UI resources and a client SDK that renders them inside sandboxed iframes. Tools point to their UI via a metadata resource URI the host fetches and displays.

CodeGraph

A local code knowledge graph for AI coding agents.

Useful forLetting a coding agent (Claude Code or another) answer architectural questions and trace flows while cutting token use and file scanning, on large codebases where grepping every turn gets expensive.

Parses a codebase with tree-sitter into a local SQLite graph of symbols, functions, classes and call relationships, then keeps it in sync incrementally as files change. It exposes the index through an MCP server so agents can query structure and trace call flows instead of repeatedly reading and grepping source. Everything runs locally, no cloud and no API keys.

FAQ

MCP questions

What is an MCP?
MCP (Model Context Protocol) is an open standard that connects Claude to outside tools and data: a database, a CRM, a file system. Instead of writing a separate integration for each tool, an MCP server exposes those capabilities so any Claude-compatible client can use them.
Skills vs MCP in Claude?
A skill is a folder of instructions and scripts that Claude loads when a task needs it: it lives inside your project and needs no separate infrastructure. An MCP is an independent server that exposes outside tools or data over a protocol, useful when you need to connect Claude to a system that lives outside the project itself. In practice, the two often combine: a skill that calls an MCP.
What is claude code mcp?
It's how you add an MCP server to Claude Code: the `claude mcp add` command (or editing the project's configuration) registers the server, and its tools become available inside the session. From there, Claude uses them like any other tool, with no extra code in the project.