Renting the Law: AI Labs Break the Research Duopoly

By Claude and Gemini with Sid Newby | July 2026

The published opinions of every American court have always been free to read and expensive to find. That gap — between reading the law and locating the one paragraph that decides your case — is the entire business model of a two-company industry. In 2026 the gap started closing, and it closed from a direction almost nobody in litigation support was watching.

The thing you were renting was never the law

Primary law is public. A court issues an opinion, the opinion enters the public domain the instant the judge signs it, and no one owns the words. Any citizen can walk into a federal courthouse, pull Erie v. Tompkins off a shelf, and read it for nothing. This has been true for as long as there have been courthouses.

So what exactly has a litigator been paying $400 a month for?

Not the text. The text is free. What Westlaw and LexisNexis sell — what they have sold for four decades — is the finding. The headnotes that summarize a holding. The citator that tells you whether the case you are about to cite got overturned last spring. The Boolean search that pulls the right precedent out of ten million opinions in under a second. The machinery that turns a public corpus nobody can search into one a busy associate can search at 11 p.m. before a filing deadline. Two companies built that machinery. They priced it like a utility, and for a long time no third option mattered.

The arrangement has taken a lot of heat over the years, most of it deserved. Public law behind private paywalls offends people, and it should. A solo practitioner who can't warrant a Westlaw seat works with a smaller law, in a real sense, than the firm across the street. But the criticism always ran into the same wall. Complaining about the paywall did nothing to build a better index. The moat was the index, and indexes are expensive to build. Righteous anger is not an index.

What changed in 2026 is that the index stopped being the moat. Foundation-model companies figured out how to read, search, cite, and sum up the whole body of American case law. No one had to build a paid front end at all. When the finding becomes something a general model does for you inside a chat window, the thing you were renting quietly loses its scarcity. And scarcity was the whole business.

One week, two organizations that agree on nothing, one move

Here is the moment that told me the ground had shifted. On May 12, 2026, two announcements landed on the same day. Thomson Reuters — parent of Westlaw, the largest paid legal research platform in the country — connected its flagship AI system to Claude through the Model Context Protocol.^[3] The Free Law Project — the nonprofit whose entire reason for existing is that paid legal research is a moral problem — connected its platform, CourtListener, to Claude through the exact same protocol on the exact same day.^[2]

Sit a moment with how strange that is. A company whose revenue depends on the law being hard to reach, and a nonprofit whose mission is that the law should be free, both looked at the same standard and concluded they had to plug into it. They agree on nothing about how law should be distributed. They agreed completely that Claude was now a place their data needed to live.

The two integrations reveal the split screen perfectly. Thomson Reuters wired in 1.9 billion Westlaw and Practical Law documents, 1.4 billion KeyCite validity signals, and a patent-pending citation ledger. Then it kept every byte behind the CoCounsel professional paywall.^[3] David Wong, the company's chief product officer, described the goal as building "the fiduciary-grade system at the center of how legal work gets done."^[3] Translation: we will meet you inside the AI tools you now live in, and we will still charge you.

CourtListener did the opposite. Every account comes with free API access, so the MCP connector is open to anyone who signs up, at no cost. And that connector carries the whole corpus: millions of federal and state opinions, PACER and RECAP dockets, the full citation network, appellate oral argument audio, judge biographies and financial disclosures.^[2] As Bob Ambrogi put it, the same technical standard now lets "a global content company extend its premium professional platform into new AI environments while also enabling a nonprofit to give anyone with an internet connection real-time access to the primary law of the United States."^[3]

Figure 1: In roughly eight weeks, a paid incumbent, a free nonprofit, and a foundation-model challenger all repositioned around the same connective layer.

I should name the obvious conflict before someone else does: Claude, the model at the center of this story, is made by Anthropic, and Claude drafts posts for this blog. I am not going to pretend that away. I would be writing the same analysis if the connector were built by Google or OpenAI, because the argument is about what MCP does to the finding, not about whose logo is on it. The interesting part was never the vendor. It was that a nonprofit and a $70-billion-revenue data company reached for the identical lever in the same news cycle.

What Perplexity actually built, and why it stings

Six weeks later, Perplexity made the subtext text. The signals had been building for weeks — by mid-June the trade press was already reporting that the company was "getting serious about legal."^[14] On June 24, 2026, at an invitation-only event in New York aimed at in-house counsel and small-firm lawyers, the company launched Computer for Counsel.^[5][8] It is not a legal research database, and Perplexity is careful to say so — the pitch is a "research, drafting and workflow layer" that sits on top of whatever databases you already have.^[5] That modesty is doing a lot of work, because a layer that sits on top of everything is precisely what threatens the companies that sell the everything.

The mechanics matter. Computer for Counsel routes each subtask across more than 20 frontier models with no single-vendor lock-in, and it reaches roughly 400 tools through MCP connectors.^[6] For primary law it leans on Midpage, which brings U.S. case law and a citator. That citator is the exact feature that made KeyCite and Shepard's indispensable. And Midpage offers it to Perplexity users free, subject to plan-based caps.^[5][10] The connector list reads like a census of the modern legal stack: Box, Carta, Clio, DeepJudge, DocuSign, Ironclad, NetDocuments, and Microsoft 365. So the agent can draft in Word, pull files from SharePoint, and read context out of Outlook and Teams.^[6][7]

Read that connector list again with a vendor's eyes. Every one of those tools used to imagine itself as a destination — the place the lawyer logs into and stays. Computer for Counsel treats each of them as a socket. The lawyer lives in the agent; the other tools become inputs the agent calls when it needs them. Joe Green, chief innovation officer at Gunderson Dettmer, praised the tool for giving lawyers "rapid, cited context" they can validate quickly.^[5] That is a compliment to Perplexity and a warning to everyone whose business assumed the lawyer would keep logging directly into their product.

Figure 2: The finding used to be a destination you paid to enter. It is becoming a function an agent performs across whichever sources you connect.

The number that explains the anxiety

Why does a $6-billion-a-year duopoly flinch at a research layer that openly refuses to build a database? Because the flinch is not about the database. It is about the seat.

Westlaw's base platform runs roughly $150 to $400 per user per month, with its CoCounsel AI features adding another $100 to $200 on top; Lexis+ lands in a similar band.^[9] Multiply that across a firm and the annual number gets serious fast. The whole reason for those seats has been that you cannot get the finding anywhere else. Then a lawyer can point Perplexity at Midpage, or open Claude with CourtListener switched on, and get a cited answer grounded in primary law. At that point the seat has to justify itself on something narrower than "we are the only door to the corpus." That was a comfortable story. It is not comfortable anymore.

Table 1: The finding is now available at several price points that did not exist two years ago. Sources as noted.

None of this means Westlaw disappears. A firm with a malpractice policy and a partner's license on the line still wants the warranted, fiduciary-grade, current-as-of-this-morning answer. Thomson Reuters is betting the whole CoCounsel rebuild on being that answer. Fine. But "the premium tier for people who must be certain" is a much smaller and more defensible market than "the only way anyone finds a case." The duopoly is being pushed up-market, and up-market is a narrower place to stand.

The half of this that actually matters

Strip away the vendor drama and there is a genuinely good thing underneath, and it deserves to be said plainly instead of buried under market analysis.

For most of my career, the practical reality of American law came down to this. A pro se litigant fighting a foreclosure, a legal aid lawyer with a caseload three times too large, a solo in a rural county — each of them worked with a diminished version of the law. The statutes were not different for them. They simply could not afford the tool that made those statutes findable. The law was equal on paper and rationed in practice. You can be furious about that and still admit that the fury never fixed it.

The CourtListener-Claude integration takes an actual swing at it. In July, Anthropic named the Free Law Project an access-to-justice partner and built CourtListener into Claude as a connector anyone can switch on for free.^[1] It is the more sympathetic bookend to a spring in which the same company was also selling Big Law a premium legal plug-in and watching the firms go all in,^[4][12] a duality that thoughtful observers flagged early as the good, the bad, and the still-unknown of putting a general model this close to legal work.^[13] Free Law Project frames the problem with unusual precision: people "can reach the law, but they still cannot trust what they are being told about it."^[2] That sentence is the whole ballgame for AI in this domain. A chatbot that confidently invents a case is worse than no tool at all — we have 1,200-plus made-up citations logged in the courts to prove it. A chatbot wired directly into CourtListener's verified primary sources is a different animal, because every answer can be run back to an opinion that actually exists.

That is the part I would not have predicted. The organizations taking the most credible swing at the affordability problem in 2026 are not the legal-tech incumbents. Those incumbents have promised to democratize the law in every keynote for fifteen years. The real movement came from a nonprofit that believes "the law belongs to everyone" and a model company that found it useful to agree.^[2] Whether Anthropic's motives are civic or commercial barely matters here. The pro se litigant with a verified citation does not care why the door opened.

Figure 3: Approximate distribution of where legal research actually happens as of mid-2026, with the free and AI-mediated slice now large enough to matter. Share figures are directional.^[9]

The catch nobody should wave away

None of this is a clean win. Pretending it is one is how legal tech ends up overpromising for the fifteenth time. Three cautions belong right next to the optimism.

Grounding reduces hallucination; it does not end it. A model wired to CourtListener will still, at the margins, misread a holding it pulled correctly. Or it will reach for a case the connector did not return and paper over the gap. The failure mode shifts from "invented a citation" to "misread a real one," which is harder to catch because the citation checks out. The check-your-cites discipline the courts have demanded since the first sanctioned brief does not relax because the tool got better. It moves up a level.

The connector economy cuts both ways on data governance. Every MCP endpoint that lets an agent reach into your DMS, your review platform, or a client's SharePoint is also a new path for privileged material to travel somewhere it should not. The same plumbing that makes Perplexity's 400 connectors powerful makes them a surface a litigation team has to map, because "the agent can reach everything" and "the agent should reach everything" are very different sentences.

And access is not the same as competence. Handing a self-represented litigant a verified citation is a real gift, and it is also the easy 20 percent of the problem. Knowing which case controls, how to distinguish an adverse one, and what to actually file is the hard 80 percent, and no connector ships that. The door being open matters enormously. It does not walk anyone through it.

What this means if you run a litigation shop

For those of us in litigation support, the temptation is to treat all of this as a legal-research story that happens next door to eDiscovery. It is not next door. It is the same house.

The Model Context Protocol is the load-bearing beam under every one of these announcements — Thomson Reuters used it, Free Law Project used it, Perplexity built its entire product around 400 of these connectors.^[3][6] MCP is now the way legal data gets wired into AI, and eDiscovery data is legal data. The same standard that lets Claude pull an opinion from CourtListener will let an agent pull from your review platform, your processing tool, your DMS. The connector economy that is dissolving the research duopoly is coming for the idea that a review platform is a place you log into. It came for Westlaw exactly the same way.

A few practical consequences worth taking seriously now, not after they arrive:

• Verification beats fluency, every time. The lesson of the CourtListener integration is that the value is not the AI's eloquence — it is the grounding in a source you can check. Any AI touching your matter should be answerable back to a document that exists. If your vendor cannot show you the chain from answer to source, that is the question to press.
• Watch what becomes a connector. When a capability turns into an MCP endpoint, its pricing power erodes, because it stops being a place and becomes a function. Legal research is the first domain to feel this. Document review will not be the last.
• Free and verified is now a real tier. For low-stakes research, early case assessment, or a client who genuinely cannot pay, "Claude plus CourtListener" is a defensible answer in 2026 in a way it was not in 2024. Knowing when the free-and-grounded tier is good enough — and when a matter demands the warranted, paid, fiduciary-grade answer — is becoming a professional skill in itself. It is a judgment call now, not a line in the budget.
• The gatekeeping was always structural. The paywall was one expression of a deeper fact: whoever controls the finding controls the practice. The finding is being redistributed. Pay attention to who ends up holding it, because the answer will not stay settled for long.

For four decades the shortest description of legal research was: the law is free, and finding it costs a fortune. In 2026 a nonprofit and a handful of model companies started making the second half of that sentence untrue. They did it not by building a better paywall or filing a better lawsuit, but by turning the finding into something a machine does for anyone who asks. The incumbents will survive by climbing up-market into certainty and warranty, which is a real and valuable place to be. But the door to the corpus, the one that stayed locked for a very long time, is standing open in a way it has not before — and a lot of people who could never afford the key are already walking through it.

Related Reading

• The People vs. PACER: How Free Law Project Is Building the Open-Source Infrastructure to Make Court Records Free
• Your Legal Research Vendor Is a Data Broker. Now What?
• The 30-Dollar Legal Agent: Microsoft Ships Inside Word