How no-code MCP works

Build absolutely anything...with the right context.

When you create an MCP server through the wizard, MCP Studio turns source content into searchable context and returns it through the Model Context Protocol (MCP).

Audience	Technical evaluators, platform teams, developers
Prerequisites	Understanding of MCP clients and access to source content you want to index

Context pipeline

1. Crawl: MCP Studio visits your sources and extracts content from web pages, GitHub files, PDFs, and federated MCP sources.
2. Index: Content is organised into passages that preserve headings and code blocks, then indexed for search. Indexing runs in the background and resumes on its own, so large sources finish reliably.
3. Search: Both keyword and meaning-based matching are used to return the passages that best answer the agent's query, favouring your current documentation.
4. Rank: When the leading passages score closely enough that the answer could be any of them, a relevance model reads the agent's question against each candidate and reorders them before the answer is assembled.
5. Deliver: The MCP runtime returns results to any connected MCP client, such as Cursor, Claude Desktop, or Windsurf.
6. Measure: Telemetry powers Core, Action, and Predictive dashboards so you can see which context creates value.

How passages are ranked

Retrieval runs in stages, and each one is there because the stage before it cannot do the job alone.

Keyword and meaning-based matching run together. Keyword matching finds exact terms, which is what you need for an identifier, a flag, or an error string. Meaning-based matching finds passages that answer a question phrased in words the documentation never uses. Keyword scoring is calculated against your server's own content, so a term that is unremarkable in one corpus and highly distinctive in another is weighted correctly for yours.

A relevance model breaks ties. The first two stages score a passage against a query independently of the other candidates, which is fast but leaves close calls unresolved. When the top passages are separated by little, a relevance model reads the question and each candidate together and reorders them. It runs on the full candidate set rather than the visible results, so a passage that was ranked too low to appear can still surface.

This stage is skipped when the leading passage is a clear winner. That is deliberate rather than a cost saving: when keyword and meaning-based matching already agree decisively, there is no tie to break, and re-deciding a settled ranking is more likely to disturb it than improve it.

Your current documentation is preferred. Where a source contains several versions, passages from the current version rank above archived ones, and an answer drawn from a version-pinned page says which release it describes.

Retrieval rules are applied as preferences. A rule you write raises or lowers how strongly a source is preferred, and can require that a source is always consulted for certain questions. Rules steer ranking; they never replace it. See Retrieval rules.

Confidence is disclosed, at two levels. If nothing retrieved is a strong match for the question, the response says so rather than presenting unrelated passages as an answer. And when your content clearly covers the subject area but does not state the specific answer, the response says that too, and asks the agent to check the passages against what was actually asked. A tidy list of nearly-right results is indistinguishable from a good answer to a connected agent, and it will answer confidently from it — so the tools tell the agent how much the passages are worth, and the agent can say it does not know instead of guessing.

Indexing

After you deploy, sources begin indexing automatically as background jobs. You do not need to keep the dashboard open. The MCP server is usable during indexing because it uses live fetching as a fallback until indexing completes.

About indexing times

• 1-3 sources: Usually 1-6 minutes, based on source size and accessibility.
• Large docs sites: Continue in the background. The dashboard shows percent complete, pages indexed, and a stalled warning if progress stops.
• GitHub repos: Often index faster for markdown-heavy content and slower for large codebases.
• Sources auto-refresh daily at 7:00 AM Central US time. Content that has not changed is skipped.
• You can manually refresh any source from the dashboard, including while it is already indexing.
• Each source supports up to 5,000 pages, and a run that fails leaves your last good index intact.

MCP protocol

Your server is reached over one endpoint:

Transport	Endpoint	Use case
Streamable HTTP	POST /api/mcp/{slug}	Every MCP client

It follows the Model Context Protocol specification, so your server works with any compliant client.

Analytics loop

Analytics help you answer product and operations questions:

• Which sources do AI agents use most often?
• Which exact passages answer user questions?
• Which content gaps prevent useful answers?
• Which sources should you improve first?