AI Agents for Lawyers

Glossary

Model Context Protocol (MCP) — Definition for Lawyers

Model Context Protocol (MCP) is the open standard that lets Claude, ChatGPT and Copilot read your DMS or docket feed through one connector. What a server exposes and what to check.

By Dr. Niklas SchmidtUpdated 13 November 20264 min read
On this page
  1. What an MCP server exposes
  2. Who supports it, as of September 2026
  3. The security questions to ask before switching one on

The Model Context Protocol (MCP) is an open standard that lets an AI model connect to outside systems, such as a document management system, a court docket feed or an e-signature platform, through one common interface instead of a bespoke integration for each. Anthropic created it in November 2024 and donated it to the Linux Foundation in December 2025; Claude, ChatGPT, Gemini and Microsoft Copilot all speak it.

What an MCP server exposes

The vocabulary is small. The client is the application running the model: Claude, ChatGPT, Copilot Studio, Cursor. The server is a program that sits in front of a data source and publishes a menu. The menu has three kinds of item: tools, which are actions the model may call (search documents, list docket entries, send for signature); resources, which are data the model may read; and prompts, which are reusable templates the server offers. The model decides when to call a tool, the client executes the call, and the server decides what the calling user is allowed to get back.

In a law firm the sequence looks like this:

  1. You ask Claude to find the executed version of a share purchase agreement in iManage and list every change-of-control trigger.
  2. Claude calls the iManage server's search tool. The call runs under your credentials, not a service account.
  3. The server returns only documents you are permitted to see. Ethical walls apply because the DMS applies them, not because the model was told about them.
  4. Claude reads the document and drafts the list with clause references.
  5. If a later step would file, send or sign, a server built to the 2026-07-28 specification can pause mid-call and ask you to confirm before it proceeds.

The document never leaves the system of record to be pasted into a chat box, and the audit trail lives in the DMS. That is the difference between an AI agent that acts inside the firm's controls and a chatbot fed with copied text.

Who supports it, as of September 2026

Anthropic donated MCP to the Linux Foundation's new Agentic AI Foundation on 9 December 2025, with AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI as platinum members and more than 10,000 public servers already published. The 28 July 2026 revision of the specification made servers stateless, added multi-round-trip confirmations and hardened the OAuth flow; the maintainers reported close to half a billion SDK downloads a month. Google's A2A protocol, which reached v1.0 on 9 April 2026 with 150-plus organisations, handles agent-to-agent communication and sits alongside MCP rather than competing with it.

Server or connectorLaunchedWhat the vendor states
iManage MCP Server14 May 2026Governed, permission-bound, fully logged; "respects existing ethical walls"; works with Claude, ChatGPT, Copilot, Harvey, Legora and custom agents
NetDocuments MCP (ndMAX)May 2026Every call runs under the user's own credentials; "documents never leave the platform"; single admin toggle
Claude for Legal connectors12 May 2026, 20-plus by SeptemberiManage, NetDocuments, DocuSign, Ironclad, Relativity, Everlaw, Box, Datasite, CoCounsel, LexisNexis, Harvey, Midpage, Trellis, Definely; a docket-watcher agent runs on CourtListener, Trellis and Everlaw
Zapier MCP2026Agent access to 9,000-plus applications; Zapier's own legal team runs intake triage and contract analysis on it
Perplexity Computer for Counsel24 June 2026400-plus MCP tools routed across 20-plus models

The iManage and NetDocuments setup guide walks through the admin steps; the Claude for Legal plugins page covers what the connectors are used for once they are on.

The security questions to ask before switching one on

The protocol is open, which means anyone can publish a server, and most people who do are not thinking about ethical walls. A June 2026 report compiling several audits found 82% of 2,614 public servers vulnerable to path traversal (Endor Labs), 43% to command injection, 53% relying on static API keys against 8.5% using OAuth, 492 servers exposed with no authentication at all (Trend Micro), and 24,008 secrets sitting in MCP configuration files on GitHub. In September 2025 a package called postmark-mcp silently blind-copied every email it sent. In a test with five servers, compromising one gave a 78.3% attack success rate across the set.

The controls are known. Anthropic's Managed Agents keep credentials in a vault so that Claude-generated code never sees a token. Its Compliance API, on the other hand, does not capture MCP server configurations, so the firm's own logging has to. Before enabling a server for anyone handling client files:

  1. Vendor-maintained, OAuth only. No community server touches client data, however useful it looks.
  2. Read-only scopes first. Add a send, file or sign tool later, and only behind the mid-call confirmation.
  3. Test the wall. Log in as a walled user and ask the agent for the walled matter. Do it before go-live and after every server update.
  4. Log where you can read it. Confirm which audit tool captures MCP calls; assume the model vendor's does not.
  5. Assume injection. A document the agent reads can carry instructions aimed at the agent, and Anthropic's own researchers say no browser agent is immune. The mechanism is explained under prompt injection and the wider controls in the guide to agent security.
  6. One toggle to turn it off. NetDocuments built this in; ask every other vendor where theirs is.

Module 4 of the AI Lab for Lawyers course covers agents and automation, the layer where MCP connectors sit.

Frequently asked questions

What is an MCP server in plain terms?

A small program that sits in front of a system such as a document management platform and describes, in a standard format, what an AI model may do with it: which searches it can run, which records it can read, which actions it can take. The model's application connects to it, asks for the menu, and calls items on that menu as needed. Because the format is standard, one server works with Claude, ChatGPT, Copilot and any other compliant client.

Is MCP safe to use with client documents?

It can be, if the server is the vendor's own and it runs under your credentials. iManage's server is permission-bound, logged and respects existing ethical walls; NetDocuments says documents never leave the platform and IT can switch it off with one toggle. The risk sits in community-built servers: a June 2026 audit found most public servers vulnerable and hundreds exposed without any authentication. Treat an MCP server as a new vendor and diligence it accordingly.

Which legal tools have MCP servers or connectors?

As of September 2026: iManage and NetDocuments document management; Claude for Legal's 20-plus connectors covering DocuSign, Ironclad, Relativity, Everlaw, Box, Datasite, Thomson Reuters CoCounsel, LexisNexis, Harvey, Midpage, Trellis and Definely; Zapier's MCP for 9,000-plus applications; and Perplexity's Computer for Counsel, which reports 400-plus MCP tools. Harvey and Legora also connect to the iManage server as clients.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion