# What the MCP server is

ResponseHub runs a **Model Context Protocol (MCP) server**. MCP is the standard
way an AI assistant connects to an outside system, so connecting ResponseHub to
an assistant lets you ask it things like "what does our policy say about data
retention?" or "start a questionnaire from this spreadsheet" and have it work
against your own knowledge base rather than guessing.

Once connected, the assistant can read and write the same material you see in
ResponseHub — your sources, your knowledge base items, your questionnaires and
your RFPs.

## The server URL

```
https://app.responsehub.io/mcp
```

You can copy it from inside ResponseHub: open **API** from your account menu,
then the **MCP** tab. That screen also carries the ready-made command for
command-line clients.

## What a connected assistant can do

- Ask questions and get answers with confidence scores and citations.
- Search the knowledge base, and create, update or approve items.
- List, create, download and delete questionnaires and RFPs.
- Upload and manage source documents.

Each of those is a *tool* the assistant chooses to call. The
[tool reference](/help/mcp/tools/) lists all 21 of them and what each one takes.

## Two ways to connect

| You are using | How you connect |
|---|---|
| Claude.ai, ChatGPT | [Add ResponseHub as a custom connector](/help/mcp/connecting-claude-and-chatgpt/) — you sign in and approve access, with no token to manage. |
| Claude Code, Cursor, the Claude API, other clients | [Create an API token](/help/mcp/connecting-claude-code-and-cursor/) and pass it as a header. |

## Assistants act as you

An assistant connected to ResponseHub is not a separate user with its own
permissions. It sees exactly what your account role lets you see, it can only
change what you could change, and every change it makes is recorded in the audit
trail against your name. If your role is read-only, a connected assistant is
read-only too.

Everything is scoped to one account. If you belong to more than one, the
connection is bound to a single account rather than reaching across them — see
[permissions and safety](/help/mcp/permissions-and-safety/).

**Deleting through an assistant:** The delete tools are real deletes — the same soft delete the web UI performs.
Nothing is destroyed straight away and support can recover a deleted record, but
an assistant asked to "tidy up" can remove work you wanted to keep. Give a
connector read-only access if it never needs to change anything.

## How this differs from the API

The [JSON API](/help/api/) and the MCP server cover much of the same ground, and
both use the same API tokens. The difference is who is calling:

- **The API** is for software you write. You choose the endpoints, handle the
  responses, and build whatever you are building.
- **The MCP server** is for an assistant to call on your behalf, in
  conversation. You describe what you want; the assistant picks the tools.

The MCP server exposes a deliberately smaller surface than the API — the things
an assistant can usefully do — so anything it does not cover is an API job.

## Next steps

- [Connect Claude.ai or ChatGPT](/help/mcp/connecting-claude-and-chatgpt/)
- [Connect Claude Code, Cursor or the API](/help/mcp/connecting-claude-code-and-cursor/)
- [Browse the tool reference](/help/mcp/tools/)