There are hundreds of MCP servers now. Most of them aren't worth adding. This is the short list of the ones that actually change how you work with Claude — filtered for setup time, token cost, and daily usefulness.
We'll cover memory, search, code tools, databases, and productivity. One quick note: if you're new to MCPs and need the basics first, start there.
Three things matter:
What it does: Keeps lists, notes, and structured records outside your context window. You paste a connector URL; Claude can then add(), search(), remove() and retrieve a standing context() — your role, current projects, preferences — across every conversation.
Why it's in this list: Setup is one URL paste. Storage is per-account a structured store with full-text search. Responses are terse (no bloat). The context() tool is the killer feature — one call loads who you are into a fresh conversation without pasting a wall of custom instructions.
Best for: Running context across sessions, Notion offload, lightweight CRM, start-my-day briefings.
Token cost: Low. Search returns matching records only; context() returns a compact summary. No full-table dumps.
What it does: Stores facts as a knowledge graph. Claude can add entities and relationships and query them later.
Why it's useful: Good for open-ended knowledge accumulation — "remember that my landlord is James, he manages the flat at X, and we have a dispute about Y." Not well-structured for retrieval at scale.
Caveats: Requires local Node.js setup. Not a hosted service. Knowledge graph grows unchecked and can get expensive to query at scale.
Token cost: Medium — depends on graph size.
Get it: npx @modelcontextprotocol/server-memory
What it does: Gives Claude live web search. Ask about news, prices, documentation — anything that changes faster than the training cutoff.
Why it's in this list: The free tier is generous. Setup is an API key + one-line JSON config. Results are clean and well-structured.
Best for: Research tasks, checking current information, getting prices, documentation lookups.
Token cost: Medium — search summaries can be verbose depending on query.
Get it: api.search.brave.com → sign up → add key to MCP config.
What it does: Lets Claude fetch and read any URL. Paste a link; Claude gets the content.
Why it's useful: Indispensable for documentation lookups, reading articles, processing web content. Much cleaner than pasting raw HTML into a message.
Caveats: Can be token-heavy if the page is long. Use it for targeted reads, not general browsing.
Token cost: Varies (page-dependent).
Get it: uvx mcp-server-fetch
What it does: Gives Claude access to GitHub repos — read files, search code, create PRs, manage issues.
Why it's in this list: If you do any coding in Claude, this is essential. Claude can read your codebase directly instead of you pasting files into messages.
Best for: Code review, PR descriptions, bug investigation, documentation generation.
Token cost: Medium-high depending on repo size. Use targeted file reads, not whole-repo dumps.
Get it: github.com/github/github-mcp-server (official, Anthropic-supported).
What it does: Lets Claude read, write, and navigate local files. Scoped to a directory you specify.
Why it's useful: Great for bulk file processing, document analysis, and working with local datasets.
Caveats: Local only — doesn't work in Claude.ai without the desktop app. Write operations need careful scoping.
Token cost: File-size dependent. Set a tight directory scope.
Get it: npx @modelcontextprotocol/server-filesystem /path/to/dir
What it does: Claude queries your database directly. Ask in plain English; it writes and runs the query.
Why it's in this list: If you have structured data in a database, this is transformative — data exploration without writing queries by hand, report generation, anomaly detection.
Caveats: Read-only mode is strongly recommended for production DBs. Claude's query isn't infallible — review before running on real data.
Token cost: Low-medium. Result sets can get large; add LIMIT clauses.
Get it: npx @modelcontextprotocol/server-db.
What it does: Claude can search Slack history, read channels, send messages, and help draft replies.
Why it's useful: For teams: search across conversations, get up to speed on a thread, draft a message with the right tone for the channel.
Caveats: Slack workspace admin may need to approve the OAuth app. Token costs can spike with large history searches — use date ranges.
Token cost: Medium. Scope your searches tightly.
Get it: modelcontextprotocol.io directory → Slack.
If you're starting from zero, here's the order that makes sense:
| Priority | MCP | Why first |
|---|---|---|
| 1 | Stash | Cheapest to add, immediate value across all sessions (context across conversations) |
| 2 | Brave Search | Web access is broadly useful; free tier covers most personal use |
| 3 | GitHub or Filesystem | Only if you use Claude for code — transformative if you do, irrelevant if not |
| 4 | Fetch | Handy for reading docs; add once you have a specific use case |
| 5 | Database connectors | Only if you have a database you want Claude to query |
The Anthropic-official Memory server is surprisingly optional given how good the alternatives are. If you're already using Stash for structured records, Memory's knowledge graph is mostly redundant for typical use cases.
Each MCP tool call burns context. If you have five MCPs installed and Claude calls all of them on a complex query, you've consumed hundreds of tokens before you've even started on the actual task. This is part of why Claude's rate limits feel inconsistent — context-heavy setups exhaust limits faster.
The discipline is: only add MCPs you actually use in most sessions. An installed-but-not-used MCP doesn't cost tokens, but it does add noise to the context about what tools are available.
Stash is free to try and takes about 90 seconds to add. Sign in with Google, copy your connector URL, paste it into Claude → Settings → Integrations. Your first tool call is usually context() — it loads your standing context into the conversation in one shot.