MCP examples

Last verified: 31 July 2026

These are patterns, not scripts. You do not call the tools yourself – you ask your client something, and it decides which tools to call. Knowing the shape of a good call helps you ask in a way that gets there.

Search, then read

The core pattern. search is cheap and returns enough to triage; get_item is the expensive, capped call, so it should be used deliberately.

"Search Remly for the incident review of last quarter's outage, then read the most relevant one in full and summarize the follow-up actions."

The client calls search with a small limit, picks the best result, and calls get_item on that id. Asking it to read everything it finds burns your 3,000-per-day get_item allowance for no gain.

Ground a code change in your own docs

Useful in Claude Code, where the model has your repository but not your written context.

"Before changing the retry logic, search Remly for our documented backoff policy and cite what you find."

With GitHub connected, this pulls in the pull request discussion and docs that explain why the current behavior exists.

Meeting prep across sources

"What do we know about the Acme renewal? Check Remly for recent context."

With Google Calendar, Jira, and Granola connected, this spans the meeting history, the open issues, and what was said on the last call – which is the case where cross-source search beats searching each tool by hand.

Compile evidence with get_context

When available, get_context returns raw cited evidence with provenance, freshness, omissions, and warnings, instead of a written answer. That makes it the right tool when you want your own model to reason over the material rather than accept someone else's summary.

"Use get_context to compile what we know about our data retention policy, then tell me where the sources disagree."

Two things to watch:

• Check the omissions field. If evidence was dropped for budget or access reasons, the answer is built on a partial picture – and that is exactly when a confident-sounding answer is most misleading.
• Raise maxTokens before raising limit. More candidate documents do not help if the token budget cannot fit them; they just get filtered out after retrieval.

What will not work

Related

• Quickstart – connecting a client
• Tool reference – parameters and defaults
• Usage limits