Skip to content

05-mcp-server-catalog.md

Latest commit

 

History

History
161 lines (108 loc) · 7.49 KB

05-mcp-server-catalog.md

File metadata and controls

161 lines (108 loc) · 7.49 KB
layout default
title MCP Server Catalog
parent Part VI — Reference
nav_order 5

MCP Server Catalog

Major Model Context Protocol (MCP) servers compatible with Claude Code, grouped by purpose. Install with:

claude mcp add <name> <transport-and-args>

Servers registered this way are stored in ~/.claude.json. Note: the mcpServers block in settings.json is silently ignored — always use claude mcp add. For installation recipes and environment setup, see [Part III, Ch. 02 — MCP integration]({{ site.baseurl }}/guide/part3-extension/02-mcp-integration/).

Official / reference implementations

Maintained by Anthropic (or forks thereof). Good starting points — well-documented, stable wire format.

Server Purpose Source Notes
filesystem Read/write local files inside a scoped path Anthropic MCP reference stdio, path-scoped; prefer over ad-hoc Bash file ops where available
fetch HTTP fetch with optional rendering Anthropic MCP reference stdio; lighter than a full browser when you only need HTML
memory Knowledge-graph memory (entities + observations + relations) Anthropic MCP reference Different from Basic Memory — check which one fits your workflow
brave-search Web search via Brave Search API Anthropic MCP reference Requires BRAVE_API_KEY; lower cost than Perplexity for simple lookups
git Git repo inspection (log, diff, blame) Anthropic MCP reference Useful when you want Claude to read history without granting shell exec

Knowledge / memory

Long-lived stores that persist across sessions. Where Claude stashes learnings.

Server Purpose Source Notes
basic-memory Semantic knowledge graph — Markdown + frontmatter + wiki-links, hybrid FTS+vector search basicmachines-co/basic-memory stdio; v0.20.2+ recommended for metadata_filters and per-query min_similarity. See [Part IV, Ch. 03]({{ site.baseurl }}/guide/part4-context-engineering/03-basic-memory-mcp/).
memory Anthropic's reference knowledge-graph impl Anthropic Lighter than Basic Memory; no FTS, no frontmatter. Pick one, don't run both.

Database / data

Direct query access against your own databases. Faster and safer than granting shell access to psql.

Server Purpose Notes
postgres PostgreSQL queries, schema introspection, table reads Multiple implementations available. On WSL / Node.js pg, a wrapper appending ?sslmode=disable to the connection URL is often required — without it, pg hangs on SSL negotiation against plaintext Postgres.
sqlite SQLite queries + schema introspection Useful for FTS5-backed app databases and on-disk caches.

Web / research

External search and content extraction. Budget-aware.

Server Purpose Notes
perplexity Web search with model routing (perplexity_ask, perplexity_search, perplexity_research, perplexity_reason) Budget-aware — set a monthly ceiling (e.g. $5/month) and use perplexity_ask for quick answers, perplexity_research for multi-source investigations.
tavily Research API — search + extract Use Tavily Extract for LinkedIn — Firecrawl explicitly blocks LinkedIn (403 we do not support this site). Tavily Extract handles it.
firecrawl Scraping, structured extraction, crawling Rich extraction and page rendering. LinkedIn-blocked — route LinkedIn URLs to Tavily instead.

Browser automation

For UI testing, screenshots, form filling, and data extraction.

Server Purpose Notes
playwright Headless browser control — navigate, click, snapshot, screenshot, network capture Preferred on WSL. Stable, no session-disconnect issues. Ships with the playwright-mcp skill in the Recommended tier.
claude-in-chrome Live Chrome extension — drives your actual browser tabs Rich feature set (tabs, console, network, screenshots). Known disconnection issues on WSL — use Playwright if you hit reconnect loops. See [Part III, Ch. 02]({{ site.baseurl }}/guide/part3-extension/02-mcp-integration/) for routing guidance.

Development tools

Server Purpose Notes
github PRs, issues, reviews, file reads, code search Preferred over gh CLI for programmatic access — typed responses, no stdout parsing.
context7 Version-specific library docs (React, Django, Next.js, Prisma, hundreds more) Upstash-hosted. Use BEFORE guessing library APIs — your training data may not reflect recent changes. Reference: the context7 plugin ships with Claude Code official plugins.

Integrations / productivity

Server Purpose Notes
claude_ai_Gmail Read + send email via Gmail OAuth Anthropic-hosted connector; OAuth flow handled by CC.
claude_ai_Google_Calendar Read + create calendar events Same OAuth model as Gmail.
claude_ai_Google_Drive Read files from Drive Complements filesystem when content lives in Drive.

Security / auditing

Server Purpose Notes
ecc-agentshield Scan ~/.claude/ configs for plaintext secrets, permission gaps, prompt-injection vectors npx-invoked; scoring output. Run monthly against your global config directory. See [Part VI, Ch. 06 — Security checklist]({{ site.baseurl }}/guide/part6-reference/06-security-checklist/).

MCP-first philosophy

Before npm install X or pip install X, check whether an MCP server already does it. MCP servers run out-of-process, expose a typed tool interface, and don't pollute your project's dependency tree.

Task MCP Preferred Don't install
Browser automation playwright Playwright/Chromium as project dep
Database queries postgres, sqlite psql/sqlite3 CLI
Web research perplexity, tavily curl/wget scripts
Knowledge storage basic-memory Ad-hoc JSON files
GitHub operations github gh CLI for read paths

In the Recommended tier this is enforced via rules/mcp/mcp-first.md.

Installing

Standard pattern (stdio transport via npx):

# Official reference impl
claude mcp add filesystem npx @modelcontextprotocol/server-filesystem ~/work

# Community / third-party
claude mcp add perplexity npx @perplexity-ai/mcp-server

# With an env var for the API key
claude mcp add tavily -e TAVILY_API_KEY=$TAVILY_API_KEY -- npx @tavily/mcp-server

Verify:

claude mcp list            # see all registered servers
claude mcp get <name>      # inspect one server's config
/doctor                    # CC 2.1.110+ warns on MCP scope conflicts

Scope precedence

When the same MCP server name is registered at multiple scopes, CC resolves in this order (closest wins):

  1. --scope=local (per-project, per-user)
  2. --scope=project (per-project, shared via .mcp.json)
  3. --scope=user (global, all projects)

Default scope is user. Override with claude mcp add --scope=project <name>.

See also

  • [Part III, Ch. 02 — MCP integration]({{ site.baseurl }}/guide/part3-extension/02-mcp-integration/) — full installation walkthrough
  • [Part IV, Ch. 03 — Basic Memory MCP]({{ site.baseurl }}/guide/part4-context-engineering/03-basic-memory-mcp/) — deep dive on the knowledge-graph server
  • [Part VI, Ch. 06 — Security checklist]({{ site.baseurl }}/guide/part6-reference/06-security-checklist/) — secrets, permissions, and MCP audit