MCP Reaches the Law Firm: IP Practice Tests the Protocol

The notice did not come from an Anthropic keynote or from a GitHub repository. It came from IPWatchdog, a publication read by patent and trademark attorneys for years, which on 5 August announced a webinar on AI customization for IP practice focused on connecting tools, data and workflows with MCP.

That detail is worth more than plenty of product announcements. When a sector with immovable legal deadlines, professional privilege and structural risk aversion starts training its professionals on an integration protocol, that protocol has stopped being a developer curiosity.

What is actually being taught

MCP, Model Context Protocol, is the open standard Anthropic published in late 2024 so a model can use external tools and data without writing a bespoke integration for every combination. An MCP server exposes three things: tools (actions the model can run), resources (data it can read) and prompts (usage templates). The client, whether Claude Desktop via claude_desktop_config.json or Claude Code, connects and discovers what is available.

The protocol adds no new intelligence, it adds a common contract: whoever maintains the firm's software publishes a server once and that server works for any compatible client, instead of repeating the integration tool by tool.

Translated into an intellectual property firm, that means wrapping in MCP servers the docketing system where deadlines live, the document management system, time recording and public patent databases such as the USPTO Patent Public Search or the EPO open services. The difference from pasting documents into a chat is substantial: the model queries the live source with the credentials you granted it, not a frozen copy in the clipboard.

The key word in the title is customization

The webinar is not called how to adopt AI. It takes the model for granted and focuses on the plumbing, which is exactly where projects break. In the Claude Code stack that plumbing has concrete names:

Skills: firm specific procedures packaged as reusable instructions, for example how a response to an office action is prepared.
Subagents: delegation to specialized agents, one for prior art search and another for deadline review.
Hooks: commands that run on lifecycle events such as PreToolUse or PostToolUse, useful as a checkpoint before writing to a critical system.
Plugins: the way to distribute all of the above to the rest of the team without everyone assembling it by hand.

Where this breaks

The risks are not theoretical. A poorly scoped MCP server over a document management system does not expose one matter: it exposes all of them. Professional privilege requires segmenting by matter, not by user. You need an exportable trail of who queried what and when. And it is wise to start read only, because letting an agent write to docketing without human review is the fastest way to turn an efficiency into a negligence claim.

None of this is exclusive to the legal sector, but there the cost of an error has a date and consequences: a missed deadline cannot be fixed with a rollback.

Who this is useful for

For midsize firms with no technical team of their own, which until now chose between whatever vertical software their vendor sold them or nothing. For in house intellectual property departments, with data spread across five systems that do not talk to each other. And for integrators, because the work here is not training anything: it is designing permissions, wrapping old APIs and deciding what does not get automated.

We have spent months building MCP servers for clients and the pattern repeats: the hard part is never the protocol, it is agreeing on what the agent may touch and with which credentials. That a sector as reluctant to experiment as intellectual property is already at the webinar stage suggests the conversation is shifting from which model to use toward who has access to what.