Skip to main content

Ask your own LLM about Measured Cloud via MCP

Status: available (self-host); hosted endpoint rolling out. This uses the MCP server, which you can run yourself today; a hosted public endpoint is being finalized. Instead of an in-app "Ask AI" chatbot, your MCP-connected LLM answers your "how do I…" questions from the product docs and cites the source — no server-side LLM and no per-answer AI spend.

Goal​

You'll ask your own MCP-connected LLM (Claude Desktop, the claude CLI, Cursor, …) a plain-English question about Measured Cloud — "how do I connect a Snowflake source?", "what's a rule group?", "why did my validation fail?" — and get an answer grounded in the official documentation, with a link to the exact doc section it drew from. The model does the reasoning; Measured Cloud's mc_search_docs tool just hands it the relevant docs to answer and cite.

This is the MCP-native replacement for an in-app Ask-AI chat. The knowledge base stays the docs; your client's model fronts it.

Prereqs​

  • A connected MCP client pointed at the MCP server, authenticated with a dqk_… personal API key (mint one). Any role works — docs search reads nothing from your account and touches no workspace data, so even a viewer-scoped key can use it.
  • That's it. mc_search_docs makes no LLM call and no network request — it searches a documentation index bundled inside the MCP server, so it works offline and needs no AI provider configured.

How it works​

The MCP server exposes a read-only tool, mc_search_docs, that ranks the product documentation against your question and returns the top few sections. Each hit carries:

FieldWhat it gives the model
titleThe doc page's title — e.g. Connect a Snowflake or BigQuery datasource.
headingThe specific section the snippet comes from — e.g. Steps.
snippetA plain-text excerpt of that section, to answer from.
urlThe public docs.measured.cloud link (deep-linked with a #anchor for sub-sections) — the citation the model gives you to read further.

The search is a small keyword scorer over an in-package index of the docs markdown — headings weigh more than body text, so "how do I connect Snowflake" ranks the Snowflake connect guide first. There is no server-side LLM: the intelligence is entirely your client's model, consistent with the primitives philosophy. When nothing matches, the tool returns an empty result with a note, so the model tells you it couldn't find it rather than inventing an answer.

Steps​

  1. Ask your model a plain-language question about Measured Cloud:

    "How do I connect a Snowflake datasource in Measured Cloud?"

    The model calls mc_search_docs with your question, reads the returned snippets, and answers — citing the url of the section(s) it used.

  2. Follow the citation. The answer links to the exact doc section (e.g. https://docs.measured.cloud/how-to/connect-a-snowflake-or-bigquery-datasource#steps). Click through when you want the full guide — the same "Read the full guide" target the in-app Help drawer links to.

  3. Refine if needed. Ask a follow-up ("what about BigQuery?", "and how do I test the connection?") — each question is a fresh mc_search_docs call, so the model stays grounded in the docs rather than drifting.

Verify​

  • The model's answer cites a real docs.measured.cloud URL, and clicking it lands on the section the answer paraphrased.
  • Asking about something the docs cover (datasources, rules, validations, alerts, SSO, self-hosting) returns a relevant page; asking about something they don't returns a "couldn't find it" note rather than a hallucinated answer.
  • Running mc_search_docs needs no LLM provider configured — it works even on an instance with no AI set up, because it does no generation itself.

Caveats​

  • Retrieval, not generation. mc_search_docs returns doc sections; your model writes the answer. A grounded snippet keeps it accurate, but review anything surprising and follow the cited link.
  • Docs are the knowledge base. The tool can only surface what's written in the docs — it doesn't see your org's data, results, or config. For those, the model uses the other MCP tools (mc_whoami, mc_datasource_list, mc_result_list, …).
  • Index freshness. The docs index ships with the server build; a self-hosted server picks up doc updates when you rebuild its image. The hosted endpoint tracks the published docs.

See also​