Before the last piece on this blog went up (an analysis of the Fifth DCA’s decision in Barfield v. Case Confidential), I ran every citation in the draft through two independent legal databases at once. An AI assistant queried my own Florida case-law system and the Free Law Project’s CourtListener, aligned the results citation by citation, flagged the one discrepancy (a missing reporter cite in CourtListener’s metadata, not an error in the draft), and produced a verification table. The whole exercise took about a minute and cost effectively nothing.
The interesting part of that story is not the AI. Language models have been able to read briefs for years, and as the profession has learned at some cost, reading is not the same as verifying. The interesting part is the plumbing. The assistant didn’t scrape websites or guess from training data. It placed structured calls to two unrelated services, operated by two unrelated organizations, that had never been integrated with each other. The calls worked because both speak a common protocol.
That protocol is called the Model Context Protocol, and if you practice law, it is worth twenty minutes of your attention. Not because you need to understand the engineering (you don’t), but because it is quietly rearranging the economics of legal research, and the rearrangement will reach your practice whether you participate in it or not.
What MCP actually is
The Model Context Protocol (MCP) is an open standard, released by Anthropic in November 2024, that defines how AI applications connect to outside systems. Those systems can be databases, document stores, court dockets, calendars, or anything with information or capabilities an AI assistant might need.
The lawyerly analogy is closer than you might expect. MCP is a citation system. The Bluebook does not contain any cases. What it contains is an agreed format (volume, reporter, page) that lets any reader locate any authority in any library without knowing in advance how that library is organized. MCP does the same thing for software. A service publishes a set of “tools” in a standard format (search these cases, fetch this statute, check these citations), and any AI application that speaks the protocol can discover and use them, with no custom integration built for that particular pairing.
Before MCP, connecting an AI system to a data source meant bespoke engineering on both ends, repeated for every combination. Five AI applications and five databases meant twenty-five integrations. With a shared protocol, it means ten. Each side implements the standard once, and every combination works. That arithmetic is the entire trick, and it is the same trick that made the web (HTTP), email (SMTP), and, yes, legal citation work.
It has already outgrown its parentage. MCP began as Anthropic’s project, but OpenAI, Google, and Microsoft adopted it within months, and in December 2025 Anthropic donated the protocol to the Agentic AI Foundation, a new home under the Linux Foundation co-founded with OpenAI and Block, with Google and Microsoft among the supporting members. Such a donation is the standard tech-industry mechanism for declaring that something has stopped being a product and become infrastructure. By early 2026, independent censuses counted over seventeen thousand publicly available MCP servers, and the protocol’s software libraries were being downloaded tens of millions of times a month. Whatever happens to any particular AI vendor, the protocol now belongs to no one. That is precisely what makes it durable.
What it’s useful for now
Concretely, today, in a law practice, it has three uses.
Research that spans sources. An AI assistant connected to multiple MCP servers can run one question across all of them in a single session, whether a commercial research platform, a free public corpus, your firm’s document management system, or a court’s docket feed. The assistant orchestrates; the sources answer. Legal technology vendors have noticed. NetDocuments built an MCP server with Anthropic so a firm’s repository can be queried where it sits, the Free Law Project runs one for CourtListener, and the major research platforms are being pulled in the same direction by their customers.
Citation verification. This is the use case I opened with, and for Florida lawyers it stops being optional on June 15, 2026, when the amendment to Rule of General Practice and Judicial Administration 2.515 takes effect. Every signature on a Florida court filing will certify that the authorities cited in it “exist and are accurately cited.” The certification is tool-agnostic. It does not care whether you used AI, an associate, or a typewriter. But it makes cheap, systematic verification suddenly very valuable. Checking a brief’s citations against two independent databases used to be an afternoon of associate time. Over MCP, it is a single instruction. When the cost of verification collapses, the excuse for skipping it collapses with it.
Your own data, finally useful. The less obvious application faces inward. A firm’s accumulated work product (briefs, memos, deposition outlines, the institutional memory currently rotting in a document management system) becomes queryable by an AI assistant through the same mechanism. The protocol doesn’t distinguish between Westlaw and your own server room. For small firms especially, this is the first realistic path to the kind of internal knowledge tooling that only the largest shops could previously afford to build.
I should disclose an interest here. I operate an MCP server. FLexlaw, my Florida legal research system, exposes its case law, statutes, citator, and citation-checking tools over the protocol, which is how the verification exercise described above was possible. I built it because the protocol made it cheap to build. That fact (a solo practitioner-built research service interoperating, as a peer, with national infrastructure) is itself the argument of this essay.
What it may become
The honest answer is that nobody knows. The 2026 protocol roadmap, with working groups on agent-to-agent communication, registry governance, and enterprise authentication, reads like infrastructure planning, which is to say boring on purpose. But three trajectories are already visible, and each deserves a lawyer’s attention.
The interface disappears. Greg Lambert at 3 Geeks and a Law Blog calls it shadow UX. The AI assistant becomes an unauthorized intermediary between the lawyer and the research platform, querying the database directly over the protocol and handing back a synthesized answer. The vendor’s carefully designed website (the search bar, the filters, the browsing experience) is never seen by a human at all. The research happened; the platform was used; nobody visited it. Vendors price by the seat, and seats are exactly what this erodes.
The bundle comes apart. Legal research subscriptions have always bundled the data (cases, statutes), the editorial layer (headnotes, key numbers, citator signals), and the interface. Protocols unbundle. The data is increasingly available from public sources such as CourtListener, state court websites, and projects like mine. The interface is evaporating. What remains defensible is the editorial layer, the human-curated treatment signals and secondary sources an agent cannot reconstruct from raw opinions. Expect pricing fights as vendors try to charge separately for “agentic access,” and expect firms to discover leverage they didn’t know they had. They can keep the editorial feeds that earn their cost and route routine lookups to public pipes. The majors are already maneuvering. Thomson Reuters rebuilt CoCounsel on an agentic architecture, and LexisNexis has embedded agent workflows into Protégé. Both are bids to own the layer where the agents live.
The access-to-justice math changes. Regular readers know this is the thread I keep pulling. The marginal cost of competent legal research is the quiet variable behind much of what ails the civil justice system; it prices out the pro se litigant and strains the small firm. Expose court data, statutes, and verification tools over an open protocol, queryable by increasingly capable and increasingly cheap AI assistants, and that marginal cost falls hard. It will not fall to zero, and raw access without judgment creates its own pathologies. Rule 2.515’s pro se trap, which I wrote about in May, is one. But the direction of the change is unambiguous, and it is the right direction.
The cautions, because there are always cautions
Three, briefly.
Confidentiality. When your AI assistant queries an MCP server, your query leaves your machine and lands on someone else’s. A research question can reveal client confidences all by itself. Ask any librarian who has fielded a “hypothetical.” Rule 4-1.6 does not have an AI exception, and the Florida Bar’s Ethics Opinion 24-1 already addresses generative AI confidentiality obligations. Know who operates every server your tools are connected to, what they log, and what their terms say. The same census that counts seventeen thousand public servers is also counting seventeen thousand operators of varying trustworthiness.
Provenance. Anyone can publish an MCP server, and the protocol’s registry and trust mechanisms are immature. The maintainers say so themselves; it is a centerpiece of the 2026 roadmap. A malicious or merely sloppy server can feed an agent bad data, and an agent that trusts its tools is only as good as its worst tool. Vetting data sources used to be the librarian’s job. It still is. The librarian just needs new questions.
Verification remains yours. The protocol makes checking cheap; it does not make the certification someone else’s. A lawyer who signs a filing under amended Rule 2.515 is not certifying that the agent’s tools returned good results. She is certifying that the authorities exist and are accurately cited, full stop. MCP-enabled verification is the reason you can no longer plead impracticality, not a delegation of the duty.
Protocols outlast products
Every legal technology cycle produces products, and most products die. Protocols are different. They are boring, they are committee-governed, and they compound. Each new participant makes every existing participant more valuable. The citation system survived every reporter, publisher, and database that ever carried it. HTTP outlived every browser that mattered when it was young.
MCP may be superseded someday (protocols do get replaced), but the thing it represents will not be un-invented. That thing is legal information systems that interoperate by default, queryable by whatever intelligence the lawyer brings, including the artificial kind. For two decades, legal research has been a story about which walled garden you rented. The protocol era’s question is different, and better. What do you actually need from the garden, now that the walls are negotiable?
Worth twenty minutes. Possibly worth a strategy.
The citation-verification workflow described in the opening is real and reproducible. It cross-checks a passage against FLexlaw’s citecheck tools and CourtListener’s citation-lookup API, both over MCP. Florida’s certification amendment is In re Amendments to Florida Rule of General Practice and Judicial Administration 2.515, effective June 15, 2026. It is discussed at length in last month’s post on this blog.