# MCP (Model Context Protocol)

MCP, the Model Context Protocol, is an open standard for connecting AI agents to outside tools, so an assistant can call a service the same way whichever client it runs in.

## In plain words

Before MCP, every AI client needed its own plug-in for every service. MCP gives one shape for both sides. A service runs an MCP server that lists its tools with names, descriptions and input schemas. An agent client, such as Claude, ChatGPT, Cursor or a custom agent, connects to the server and can call those tools.

A hosted MCP server runs over HTTP at a URL. You add the URL to your client and sign in, and the agent sees the tools.

## An example

A founder asks their assistant: "Send our standard mutual NDA to the two people on tomorrow's calls." The assistant calls `list_templates`, picks the NDA, calls `create_from_template` twice and asks to send both. The founder gets two approval links, reads each draft and approves.

## How it shows up in docuinked

docuinked runs a hosted MCP server at `https://docuinked.com/mcp`. Clients that can send a header use a Bearer API key. Clients that sign in by browser use one-click OAuth, where you approve the connection yourself. Setup lines for each client are on the [MCP docs page](/docs/mcp).

The MCP tools are the same as the REST API, from one registry. Agents can create drafts, list templates, read status and fetch evidence. They can never sign, and there is no signing tool. Over MCP, every caller counts as an agent, so a live send never goes straight out. It returns an `approval_url` for a named person, bound to the `draft_hash`. See [human-in-the-loop](/glossary/human-in-the-loop) and [agents](/agents).

## Common confusions

**MCP is not a security model.** The protocol carries tool calls. Whether a dangerous call needs a human is up to the client or the server. The MCP spec says clients should let a person deny tool calls, but custom agents often have no such client. docuinked enforces approval on the server instead.

**Having an MCP server is common now.** Several e-signature vendors run one. What matters is what each server allows an agent to do.

## Related terms

- [Human-in-the-loop](/glossary/human-in-the-loop)
- [Idempotency](/glossary/idempotency)
- [Template](/glossary/template)

## Sources

Checked 2026-11-07.

- [MCP tools specification](https://modelcontextprotocol.io/specification/2026-07-28/server/tools)
- [Cipher Projects on e-signature MCP servers](https://www.cipherprojects.com/blog/posts/esignature-mcp-servers-ai-agents-2026/)

## Common questions

### Is MCP only for one AI company's products?

No. It is an open protocol. Many clients support it, including desktop assistants, coding tools and agent frameworks such as n8n's MCP Client Tool.

### What is the difference between an MCP tool and an API endpoint?

An MCP tool is described for a model to read, with a name, a plain description and a schema. In docuinked each tool is also a REST route, so both paths do the same work.

### Can an MCP server see my other conversations?

No. A server only receives the tool calls the client sends it, with their inputs. It has no view of the rest of the chat.