Claude Legal Research MCP: How to Use Claude for US Legal Research

To use Claude for legal research, connect it to a legal research MCP server. MCP (Model Context Protocol) lets Claude call live tools instead of guessing from training data, so it searches real US court opinions and returns citations you can open and check. You add the server once (a connector or a short JSON config), then ask Claude legal questions in plain English. This guide shows the exact config and a real query, end to end.

Anthropic made this the default story for law on May 12, 2026, when it launched Claude for the legal industry with 20+ MCP connectors and 12 practice-area plugins (Anthropic, May 2026). Legal research is one of the headline use cases. The rest of this post is how to wire it up and what it looks like once it works.

Why Claude Hallucinates Citations Without MCP

Ask Claude about the Fourth Amendment and warrantless cell-site searches and you'll get a clean textbook answer. Ask it to find the exact citation for Carpenter v. United States, with the parallel reporters and the paragraph where the Court held that accessing historical CSLI is a search, and it falls apart.

It will hallucinate a citation, mangle the parallel cites (the real ones are 585 U.S. 296, 138 S. Ct. 2206, and 201 L. Ed. 2d 507), or hand you a quote that isn't in the opinion.

This isn't Claude's fault. LLMs are trained on general web data. They don't have access to structured legal databases, and they certainly don't have access to millions of US court opinions with verified citations and treatment histories.

But what if they could?

A Claude legal research MCP fixes this. You give Claude (or Cursor, or VS Code Copilot) an MCP server that exposes US legal-research tools, and you can use Claude for legal research without crossing your fingers on every cite.

When you ask about US case law, Claude doesn't guess. It searches real opinions, pulls real citations, and grounds its answer in actual court records.

One quick clarification before we go further. The public REST API and the MCP server are statutes and legislation only (U.S. Code, CFR, 49 state codes, the U.S. Constitution, Federal Rules, and Executive Orders).

Case-law search, the ask-a-question surface, citation resolution, and treatment lookups are features of the Vaquill AI product and app, not public REST or MCP endpoints.

I'll walk you through the setup and show you what this looks like in practice. Five minutes, and you'll have Claude doing legal research for you. No coding required.

If you're weighing the AI research stack as a whole, we wrote up how this replaces a Harvey-style setup with Claude plus an MCP server.

TL;DR

• Use Claude for legal research by connecting a legal research MCP server. LLMs invent case citations because they have no real legal database behind them. MCP gives Claude live tools that read from one.
• Since the May 12, 2026 Claude for Legal launch, you can add many legal MCP connectors from Claude's built-in directory; for a self-hosted server, you edit one JSON config block (about two minutes).
• Connect the MCP server once and Claude can search 8M+ US federal and state opinions on demand.
• You get eight tools: ask a question, search case law, quick search, resolve a citation, search by citation, look up treatment history, traverse a citation network, and check pricing.
• Every answer comes back with citations you can open and verify. No hallucinated reporters.
• Works in Claude Desktop, Claude Code, Cursor, VS Code (Copilot), and Windsurf.
• Case-law research and AI-grounded answers are features of the Vaquill AI app; the statute API and MCP server cover the U.S. Code, CFR, and 49 state codes.

Part of our MCP and developer guide series.

For related MCP / API / developer coverage, see Legal Research from Inside Cursor and Claude Code: The MCP Way and Adding Legal Research to Your AI Agent with MCP.

Why Use Claude for Legal Research (and Where It Breaks)

Claude is a strong reasoning engine, but on its own it has no idea whether the citation it just wrote is real. The fix is not a better prompt. It is giving Claude a tool that reads from an actual legal database.

The rest of this post is how to do that, and what it looks like once it works.

What is MCP?

MCP stands for Model Context Protocol. Anthropic released the spec in late 2024, and it has since been adopted across most major AI coding tools.

Think of MCP as a plugin system for LLMs. Instead of an AI assistant being limited to its training data and whatever you paste into the chat window, MCP lets it call external tools, fetch live data, and take actions on your behalf.

The AI sees a list of available tools with descriptions, decides which ones to call based on your question, and folds the results into its response.

It's the difference between an AI that knows about legal research and an AI that can actually do legal research.

Here is the loop, end to end. Claude reads your question, picks a tool, the MCP server queries a real legal database, and the citations flow back into the answer before Claude ever writes a word:

Two Ways to Connect a Legal Research MCP to Claude

There are two paths, and the launch in May 2026 added the first one.

1. The Claude connector directory (point and click). On May 12, 2026, Anthropic launched Claude for the legal industry with more than 20 MCP connectors, including legal-research providers (Anthropic, May 2026). To add one, open Claude on web or desktop, go to Customize then Connectors then Browse Connectors, pick the provider, and authorize it. No config file. Some connectors need a paid account with that provider (for example, the CoCounsel Legal connector needs a Thomson Reuters subscription).

2. A self-hosted MCP server (one JSON edit). Any MCP server, listed in the directory or not, can be wired into Claude Desktop, Claude Code, Cursor, VS Code, or Windsurf with a short config block. This is the path the rest of this guide uses, because it works the same in every client and you control the API key.

Which Legal Research MCP Connector Should You Use?

The May 12, 2026 launch put several legal research providers into Claude at once. They are not interchangeable. Some hand Claude raw API access to a database, some ride on a paid subscription you already have, and some return a synthesized answer with citations already attached. Pick by what you already pay for and what you want back.

The honest read: if you want free primary-source access and don't mind a raw API surface, CourtListener is the default. If your firm already runs on Westlaw, CoCounsel is the least friction. If you want case law and statutes together with a synthesized, cited answer instead of a result list, that is where a server like Vaquill AI fits. There is no reason to pick only one; MCP lets you attach several and let Claude choose.

What You Get: Eight Research Tools

When you connect the MCP server, your AI assistant gets access to eight tools. You don't need to memorize these. Claude picks the right tool automatically based on how you phrase your question.

But it helps to know what's available so you can ask better questions.

Ask a legal question (in the Vaquill AI app): Ask anything about US law in plain language. The Vaquill AI app searches actual court opinions and returns a sourced answer with citations and links to the relevant pages.

You can get a quick answer (standard mode) or a thorough, multi-step analysis (deep mode). (This is a feature of the Vaquill AI app, not a public REST or MCP endpoint.)

Search case law (search_legal_cases): Search for cases using specific terms, statutory provisions, or legal concepts. Set countryCode: "US" and filter by courtType (supreme_court, circuit_court, district_court).

Supports Boolean operators (AND, OR, NOT), proximity searches, and wildcards. Use this when you know what you're looking for and want precise results.

Quick search (quick_search): A faster, lighter search that returns just the top few results. Good for quick lookups when you don't need an exhaustive list.

Resolve a citation (resolve_citation): Give it any US reporter format (U.S., S. Ct., L. Ed. / L. Ed. 2d, F.2d / F.3d, F. Supp.) or even a party name, and get the full case record: title, court, decision date, judges, and all known citation formats.

Search by citation (search_cases_by_citation): Find cases when you only remember part of the citation or an approximate party name. More forgiving than resolve, which expects a reasonably complete citation.

Look up a case (lookup_case): Get complete details for a specific case (it also accepts a clusterId for US opinions), including treatment statistics: how many times subsequent courts have followed it, distinguished it, overruled it, or approved it.

Citation network (get_citation_network): See which cases cite a given opinion and which cases it relies on. Traverse the network up to three levels deep in either direction.

Check pricing (get_pricing): See current API pricing shown in your dashboard. This one's just for reference.

The database covers 8M+ federal and state opinions across the U.S. Supreme Court, the U.S. Courts of Appeals, and the U.S. District Courts.

For more on how Vaquill AI's research tools work and how to query them, see our MCP and developer guide series.

Setup: Claude Desktop

The entire setup takes about two minutes. You need two things: an API key and a config file edit.

Step 1: Get Your API Key

Sign up at app.vaquill.ai if you haven't already. Navigate to your API settings and generate a key. It will start with vq_key_.

Vaquill AI is self-serve; sign up to start, and current pricing is shown in your dashboard.

Step 2: Edit the Config File

Claude Desktop reads its MCP configuration from a JSON file. On macOS, the path is:

~/Library/Application Support/Claude/claude_desktop_config.json

On Windows:

%APPDATA%\Claude\claude_desktop_config.json

Open (or create) that file and add the server entry:

{
  "mcpServers": {
    "vaquill": {
      "command": "uvx",
      "args": ["vaquill-mcp"],
      "env": {
        "VAQUILL_API_KEY": "vq_key_your_key_here"
      }
    }
  }
}

Step 3: Restart Claude Desktop

Quit and reopen Claude Desktop. You should see a small hammer icon in the chat input area. Click it, and you'll see the eight tools listed. That's it. You're connected.

Setup: Claude Code (Terminal)

If you use Claude Code in the terminal, setup is a single command:

claude mcp add vaquill -- uvx vaquill-mcp

Then set your API key as an environment variable:

export VAQUILL_API_KEY=vq_key_your_key_here

Add the export line to your .bashrc or .zshrc so it persists across sessions.

Create an API keyBook a demo

Setup: Cursor

Cursor supports MCP natively. Go to Cursor Settings > MCP Servers and add this configuration:

{
  "mcpServers": {
    "vaquill": {
      "command": "uvx",
      "args": ["vaquill-mcp"],
      "env": {
        "VAQUILL_API_KEY": "vq_key_your_key_here"
      }
    }
  }
}

Restart Cursor, and the tools will be available in your AI chat panel.

Setup: VS Code (GitHub Copilot)

For VS Code with GitHub Copilot, add the config to your workspace or user settings:

{
  "mcp": {
    "servers": {
      "vaquill": {
        "command": "uvx",
        "args": ["vaquill-mcp"],
        "env": {
          "VAQUILL_API_KEY": "vq_key_your_key_here"
        }
      }
    }
  }
}

You can place this in .vscode/settings.json for per-project configuration, or in your global VS Code settings.

Setup: Windsurf

Add the following to ~/.windsurf/settings.json:

{
  "mcpServers": {
    "vaquill": {
      "command": "uvx",
      "args": ["vaquill-mcp"],
      "env": {
        "VAQUILL_API_KEY": "vq_key_your_key_here"
      }
    }
  }
}

Real Workflows: How Lawyers and Researchers Use This

Once the MCP server is connected, you just talk to Claude normally. It figures out which tools to call on its own.

Here are the workflows I see people use most, modeled on actual lawyer research patterns.

Workflow 1: Researching a Doctrine

You're briefing a suppression motion and need the current law on warrantless searches under the Fourth Amendment, the leading cases, and the tests they established.

You type:

What is the current legal standard for warrantless searches under the Fourth Amendment? I need the leading Supreme Court cases and the tests they established.

Claude calls ask_legal_question with deep mode. You get a structured answer tracing the doctrine through Katz v. United States, 389 U.S. 347 (1967), Terry v. Ohio, 392 U.S. 1 (1968), and Carpenter v. United States, 585 U.S. 296 (2018). Each principle is tied to a specific opinion with a citation.

Follow up:

Search circuit court opinions for "Fourth Amendment AND warrantless search" since 2020.

Claude calls search_legal_cases with countryCode: "US", courtType: "circuit_court", and the Boolean query. You get a list of recent appellate decisions applying the rule.

Workflow 2: Is It Still Good Law?

You're about to lean on Chevron deference in a brief. Before you do, you need to know whether it survived.

You type:

Is Chevron U.S.A. v. NRDC, 467 U.S. 837 (1984) still good law? Has any later decision questioned or overruled it?

Claude calls lookup_case and checks the treatment statistics. The overruled count is not zero. Chevron was overruled by Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), which scrapped the deference framework for agency interpretations of ambiguous statutes.

You see the followed, distinguished, and overruled counts at a glance, so you don't cite dead doctrine as if it were live.

This single check is the difference between a credible brief and a sanctions exposure.

We cover that risk in detail in our piece on AI hallucinations and legal research sanctions.

Workflow 3: Statute Lookup

You're advising on a civil rights claim and want the text of the statute plus the standard of review for a related agency action.

You type:

Pull 42 U.S.C. 1983 for me, then show me 5 U.S.C. 706 on the scope of APA review.

Claude calls search_us_statutes to locate the provisions, then get_us_statute_section to return the section text for each. You get the operative language of the civil rights cause of action and the APA review standard (which ties directly back to Loper Bright after the fall of Chevron).

Then:

Now search 17 U.S.C. 107 on fair use, and find cases on "fair use NOT parody."

Claude pulls the fair use factors from the statute and runs search_legal_cases with the Boolean query so you get fair use opinions that aren't about parody.

Other provisions you'll reach for the same way: 28 U.S.C. 1331 (federal-question jurisdiction) and 18 U.S.C. 922 (firearms).

Workflow 4: Writing a Legal Memo

You need to produce a memo on qualified immunity in an excessive-force case brought under Section 1983.

You type:

I need to write a memo on qualified immunity for an excessive force claim under 42 U.S.C. 1983. What are the key Supreme Court decisions and the two-step test? Search "qualified immunity AND excessive force" and include the specific holdings.

Claude uses ask_legal_question in deep mode plus search_legal_cases with the Boolean query to compile a thorough answer covering the clearly-established-law standard and the leading authorities. Each holding is cited with its source.

You can then say:

Now draft the memo. Structure it with an introduction, the legal framework, the key cases in chronological order, the current position, and a conclusion.

Claude produces a structured memo where every assertion is backed by a citation from the research it just completed.

Workflow 5: Quick Citation Lookup

You're in a meeting and someone mentions a case. You need the details immediately.

You type:

Resolve 384 U.S. 436 for me.

Claude calls resolve_citation and returns the full case record: Miranda v. Arizona, decided in 1966, with the court, the judges, and all known parallel citations across reporters. Ten seconds, no browser tabs. The same works for Gideon v. Wainwright, 372 U.S. 335 (1963) and Mapp v. Ohio, 367 U.S. 643 (1961).

Workflow 6: Comparing and Tracing Landmarks

You type:

Compare how Brown v. Board of Education, 347 U.S. 483 (1954) and Citizens United v. FEC, 558 U.S. 310 (2010) have been treated. Which one has been cited more, and by whom?

Claude uses lookup_case for treatment stats and get_citation_network to map downstream citations. You can run the same trace on Obergefell v. Hodges, 576 U.S. 644 (2015) or go all the way back to Marbury v. Madison, 5 U.S. (1 Cranch) 137 (1803). The answer is structured, sourced, and ready to drop into a brief or memo.

Tips for Getting the Best Results

A few things I've learned from using this daily:

Be specific about courts and dates. "Ninth Circuit opinions on qualified immunity after 2020" produces much better results than "cases about police liability." The more constraints you give, the more targeted the search.

Use the right phrasing for the right tool. When you say "search for cases about X", Claude uses the Boolean search tool. When you ask "what is the legal standard for X?", it uses the AI question tool.

Both are useful. The first gives you a list of cases; the second gives you a synthesized answer with citations. You don't need to name the tool; just phrase your question naturally.

Ask follow-up questions. Claude remembers context within a conversation. If you search for a case and then ask "what did the dissent say?", it knows which case you're referring to. Build on previous results instead of starting fresh each time.

Chain your research. Start with a broad question, then drill into specific cases, then check their citation treatment.

A typical workflow might be: ask a legal question, pick the most relevant opinion from the sources, look up its citation network, then read the cases that distinguished it. All without leaving the chat window.

Check treatment before citing. Always verify that a case hasn't been overruled before relying on it. Chevron is the cautionary tale here: cite it today and you're citing law that Loper Bright erased.

Use deep mode for complex questions. If your question involves multiple legal issues, conflicting precedents, or requires tracing a line of authority, ask Claude to use deep mode.

It takes longer but produces significantly more thorough results.

FAQ

Can Claude do legal research? Yes, but not reliably on its own. On its own Claude reasons well and can hallucinate citations, because it has no live legal database. Connect a legal research MCP server and Claude searches real US opinions and returns citations you can open and verify. Always check treatment before you cite.

What is an MCP connector for legal research? MCP (Model Context Protocol) is an open standard Anthropic released in late 2024 that lets Claude call external tools. A legal research MCP connector exposes case-law search, citation resolution, and treatment lookups to Claude so it queries a real database instead of guessing. See the spec.

How do I connect Claude to a legal research database? Two ways. Add a connector from Claude's built-in directory (Customize then Connectors then Browse Connectors), which Anthropic expanded for legal on May 12, 2026. Or wire in a self-hosted MCP server with a short JSON config that works across Claude Desktop, Claude Code, Cursor, VS Code, and Windsurf. This guide shows the config for both desktop and terminal.

Is Claude for Legal a real product? Yes. Anthropic launched Claude for the legal industry on May 12, 2026 with more than 20 MCP connectors and 12 practice-area plugins (Anthropic, May 2026). Legal research is one of the headline use cases.

Does Claude have access to Westlaw or case law by default? No. Out of the box Claude has no access to Westlaw, Lexis, or court opinions. Access comes from a connector. Thomson Reuters connects CoCounsel Legal (Westlaw and Practical Law) to Claude, announced May 12, 2026 (LawSites, May 12, 2026).

Will Claude still hallucinate citations with an MCP connected? The risk drops sharply because answers are grounded in retrieved opinions with real citations, but you still verify. Ask Claude to look up the case and check its treatment statistics before you rely on it, since a real case can still be overruled.

Is it free to use Claude for legal research? Claude itself needs a paid plan for serious work. The MCP server may be free or metered depending on the provider. Vaquill AI is self-serve; sign up to start, and current pricing is shown in your dashboard.

Is the CourtListener connector free inside Claude? Yes. Free Law Project made CourtListener available in Claude's directory on May 12, 2026, and every CourtListener account includes free API access; elevated access is available through an FLP membership (LawSites, May 2026).

Does CoCounsel Legal require a paid subscription? Yes. The CoCounsel Legal connector brings Thomson Reuters data (roughly 1.9 billion Westlaw and Practical Law documents) into Claude, but you need an existing Thomson Reuters subscription to use it; pricing is quote-based, not public (LawSites, May 2026).

Which AI clients support legal research MCP servers? Claude Desktop, Claude Code, Cursor, VS Code (GitHub Copilot), and Windsurf all support MCP, so the same legal research server works in any of them with a per-client config.

For Developers

If you're building a legal tech product or an AI agent that needs to do legal research programmatically, this MCP server is the same one you'd use in your code.

For a deeper look at agent architectures, multi-step tool chaining patterns, and integration with frameworks like LangGraph and CrewAI, see our post on adding legal research to AI agents with MCP.

The public REST API and the MCP server cover statutes and legislation only (U.S. Code, CFR, 49 state codes); case-law search and the AI Q&A surface are features of the Vaquill AI app, not REST or MCP endpoints. To get started, sign up at app.vaquill.ai or email contact@vaquill.ai.

You don't need to build your own RAG pipeline to use them.

For more on plugging legal research into an AI assistant via MCP, see /features/legal-research.

Priyansh Khodiyar

Co-Founder & CTO

Priyansh leads engineering and AI at Vaquill, from the matter workbench to drafting, document comparison, document matrix, and citation-verified research.