Adding MCP servers: Context7 & Polygon
From theory to real tools. Equip Claude Code with up-to-date docs and live market data, learn the magic word that forces a server, and see why servers barely touch your context now.
Enough theory — let's equip Claude Code with real tools. We'll add Context7 (up-to-date library docs) and Polygon/Massive (live stock prices), use them, and then cleanly remove them.
01Add Context7 — always-current docs
Claude's knowledge has a training cutoff, so it can assume out-of-date APIs. Context7 lets the model ask "brief me on this library" and get back the latest documentation — a genuinely useful everyday server.
You add a server from the command line (outside Claude), then relaunch. The command comes straight from the server's GitHub or Glama page:
claude mcp add context7 -- npx -y @upstash/context7-mcp # added to local config. now relaunch: claude
Read that command like a sentence: claude mcp add registers a server named
context7; everything after -- is how to launch it — here an
npx (Node) command pulling @upstash/context7-mcp. An API key
is optional (only for a better rate limit). Run /context and you'll see two new tools
appear — resolve-library-id and query-docs — using ~1,000
tokens between them.
Whether Claude actually reaches for an MCP server can be hit-or-miss. To force it, name the server with the word use. It's not always required — often it figures things out — but it's the reliable lever.
# forcing the server explicitly: Use context7 to summarise the right way to use the OpenAI Agents SDK with a model other than OpenAI's. # often you don't even need "use": How do I turn off telemetry with CrewAI?
Watch it work: it calls resolve-library-id to find the library, then
query-docs to pull current documentation, and answers with exactly the right steps.
The second question worked without "use" — it had the handle by then. And remember what's amazing here:
not that the tools exist, but that we plugged them in so easily.
02Add Polygon (Massive) — live market data
To show the breadth of what's possible — nothing to do with coding — add a stock-price server. Polygon (now renamed Massive) needs a free API key (free tier = slightly stale data).
claude mcp add massive -e MASSIVE_API_KEY=<your-key> -- npx -y @polygon/mcp claude # then just ask: What's the current share price of Apple?
No "use" needed — it picks the right tool (get-snapshot-ticker), contacts Polygon
in the cloud (running locally, calling out remotely — remember confusion #1), and returns a price. The server
exposes a huge set of tools, but modern Claude Code doesn't load them all into context…
Run /context and you'll see a long list of Massive tools in dark grey
(available) while only the one it needs is loaded (~156 tokens). Until recently Claude Code loaded all
of them. This lazy-loading is why MCP servers are far less of a context hog than they used to be — though heavy
use still adds up.
03List and remove servers
Housekeeping is symmetrical with adding. Inside Claude, /mcp shows connected servers;
from the shell, claude mcp remove unplugs them.
# inside Claude — see what's connected /mcp # from the shell — remove them claude mcp remove context7 claude mcp remove massive # relaunch, /mcp → "no MCP servers configured"
MCP shines for very granular, specialised capabilities that need real tool signatures — live market data is the classic case. For most everyday needs you'll reach for skills or plugins instead (next). Favour skills over MCP except where MCP is the only option, like Massive.
✓ Key takeaways
- Add servers from the shell:
claude mcp add <name> -- <launch command>, then relaunch. - Everything after
--is how the server launches (e.g. annpxpackage). - Say "use <server>" to force it; often Claude picks the tool on its own.
- Modern Claude Code shows tools as available (dark grey) and only loads what it needs — big context saving.
- Manage with
/mcp(list) andclaude mcp remove <name>(unplug).