Contact · soon
All work
01 / Model Context Protocol server

Stacks Legal MCP

A 45-tool research server that lets Claude search, read, and cite New York and federal law, with every answer tied to a source a lawyer can check.

Client
Lofted Technologies product, in use at a New York law firm
Role
Architecture, data engineering, and build
Year
2026
Status
In production · private
Live
Private deployment
Claude · Stacks Legal MCP
Is Pell still the standard for arbitrary-and-capricious review? Pin cite, and how often the Court of Appeals still relies on it.
citing_references("34 NY2d 222") ✓ 0.2s
get_case("34 N.Y.2d 222") ✓
Yes. Matter of Pell v. Bd. of Educ., 34 N.Y.2d 222, 231 (1974) (determination confirmed): “Arbitrary action is without sound basis in reason and is generally taken without regard to the facts.” The citator finds 5,202 citing decisions, 247 of them from the Court of Appeals, and no formal negative treatment. Most recent Court of Appeals use: Matter of Mantilla v. N.Y.C. Dep’t of Hous. Pres. & Dev., 45 N.Y.3d 581 (2025).

45 tools

  • NY reported decisions
  • NY slip & motion decisions
  • Federal courts
  • NY Consolidated Laws
  • NYS bills
  • U.S. Code
  • Code of Federal Regulations
  • NYCRR
  • Federal Register
  • Regulatory dockets
  • Agency opinions
  • Court dockets
  • Verdicts
  • Practice treatises
  • Practice forms
  • NYS open data
  • Citator
Citator graph of Sept. 24, 2026
45research tools exposed to Claude
5M+court decisions held in full text
40M+citation links in the citator graph
800+automated tests

The brief

Why it needed to exist.

General-purpose AI models invent citations. Lawyers need a research tool that answers with real, checkable authority and says plainly when it can't find something.

The production

What we built.

The engine behind this act. It connects Claude to primary law through the Model Context Protocol: case law, statutes, bills, regulations, dockets, agency opinions, verdicts, treatises, practice forms, and state open data. Every result carries the case name, court, date, official and parallel citations, and a link back to the source.

Where public legal APIs stop, it keeps its own full-text collections of New York and federal decisions, updated nightly, and a citator built from the decisions' own text.

  1. 01Forty-five tools

    Search, retrieval, and citation tools across case law, statutes, bills, regulations, dockets, agency opinions, verdicts, treatises, practice forms, and open data, drawing on live public legal APIs and its own collections.

  2. 02A citator built from scratch

    A citation graph extracted from the decisions' own text: who cites a case (with the citing sentence), negative treatment, and direct history. Negative-treatment detection was checked by hand, and the citator is always labeled as not Shepard's or KeyCite.

  3. 03Dispositions up front

    Search results report how each case came out, so a good quotation from a losing case is visible as one before anyone relies on it.

  4. 04Connects like any Claude connector

    Standards-based OAuth 2.1 sign-in: a colleague adds it to Claude with a URL and an account, with no local setup.

  5. 05Keyword search you can check

    Full-text indexes ranked by relevance, with no embeddings, so every hit is a literal match a lawyer can verify on the page.

  6. 06Current without supervision

    New decisions and citator links are added nightly. The server respects upstream rate limits, caches what it can, and identifies itself honestly.

Under the hood

How it’s made.

  • RuntimePython with the MCP Python SDK; local and remote transports
  • DataSQLite full-text search ranked by BM25
  • CitatorCitation extraction with eyecite and a custom canonicalizer; incremental nightly updates
  • AccessOAuth 2.1 with PKCE and dynamic client registration for claude.ai
  • Quality800+ automated tests and guarded, repeatable deploys

Stack: Python · Model Context Protocol · SQLite full-text search · OAuth 2.1

Next productionStacks Cite Check for Word