See the complete route index at https://crusible.dev/llms.txt before following links.

# MCP

Model Context Protocol is one of the two ways a tool becomes the provider behind a surface. Here is what an MCP connection gives you.

<!-- https://crusible.dev/docs/reference/providers/mcp/ · doctype: reference · role: both · reviewed: 2026-09-05 -->


Model Context Protocol (MCP) is one of the two ways an external tool becomes
the <a class="docs-term" href="/docs/reference/glossary/#provider">provider</a> behind a
<a class="docs-term" href="/docs/reference/glossary/#surface">surface</a>. The other is the tool's own
API. Both are <a class="docs-term" href="/docs/reference/glossary/#connected">connected</a> mode: the
tool keeps its data, and the surface renders it.

## What the two connection routes have in common

Neither route moves anything. The tool that held the data before the
connection holds it afterwards, and it is the tool that answers for it. A
connection changes one surface. The <a class="docs-term" href="/docs/reference/glossary/#channel">channel</a>,
its <a class="docs-term" href="/docs/reference/glossary/#conversation">conversation</a>, its
<a class="docs-term" href="/docs/reference/glossary/#agent">agent</a> roster and its permission model are
identical either way.

Both routes are also bounded by the permissions granted at connection time. An
agent operating a connected tool acts as that connection, and it can do
nothing the connection was not given.

The route a tool uses changes none of the answers a reader comes here for.

| Question                    | The answer on either route                                                       |
| --------------------------- | -------------------------------------------------------------------------------- |
| Who holds the data          | The tool. Connecting moves nothing                                               |
| Who answers for it          | The tool, under its own controls                                                 |
| What an agent may do        | Only what the connection was granted                                             |
| What changes in the channel | One surface. The conversation, the agent roster and the permissions are the same |
| Where the route is named    | The provider page's **MCP or API** row                                           |

## Which route a tool uses

The route is a property of the tool rather than a choice. Each provider page
names the route in its **MCP or API** row, above that page's first heading,
and states what an agent may read and write through that connection — which is
where the route makes a difference to a reader. Nothing about the record
depends on the answer: an MCP provider and an API provider both hold their own
data, and both are named in that page's **Who holds the record** row.

## What is out of scope here

These docs describe what a connection lets you do. How the protocol itself
works, and which tools are cheap to support, are engineering questions and are
not answered on this page or on any provider page.

