Connect DocketDrift to Claude

Give an AI assistant the ability to look up real appellate opinions — and quote the court's own words instead of inventing a citation.

DocketDrift runs a Model Context Protocol server: a standard way for an AI assistant to call out to a real data source. Connect it and Claude can search opinions, pull a case by docket number, resolve a citation, and read a judge's record — all against the actual corpus, all verbatim.

No account, no API key, no sign-up. The server is public and read-only.

The address

Everything below uses one URL:

https://docketdrift.com/mcp

Claude.ai and Claude Desktop

Open Settings → Connectors → Add custom connector, paste the URL above, and save. Leave the authentication fields empty — there is nothing to authenticate.

Claude Code

From your project directory:

claude mcp add --transport http docketdrift https://docketdrift.com/mcp

Or commit a .mcp.json at the repo root so the whole team picks it up:

{
  "mcpServers": {
    "docketdrift": {
      "type": "http",
      "url": "https://docketdrift.com/mcp"
    }
  }
}

Anything else

It is a standard streamable-HTTP MCP server, so any compliant client works. Point it at the URL; it speaks initialize, tools/list, and tools/call.

What it can do

Six tools, all read-only. None of them can write, change, or delete anything — the server has no write path at all.

• Search opinions — keyword search across one state's corpus.
• Get opinion — a case by docket number, with its disposition, panel, and full text.
• Look up citation — paste a reporter cite or docket and land on the case.
• Get judge — a judge's record: how they vote, who they sit with, where they split.
• Get citing opinions — what later cases said about this one, in their own words.
• Get statute — the opinions construing a given statute.

Coverage is Minnesota, New Hampshire, Arizona, and Louisiana — Supreme Courts and Courts of Appeal, roughly 469,000 opinions, the oldest from 1809. Known coverage gaps are documented; please read them before relying on completeness.

What it will not do

The same posture as the rest of the site, and it is the whole point of connecting this rather than asking an AI to recall case law from memory:

• Nothing is generated. Every field is either verbatim court text or a fact extracted from it. Holdings quote the court's own sentence. There is no summarization step for a model to get creative in.
• Nothing is logged. Tool arguments — the thing being researched — travel in the request body, which our access log does not record, and we do not write them down anywhere else. See the privacy statement.
• Nothing is remembered. The server is stateless. There are no accounts and no sessions, so there is no identifier tying one lookup to the next.

One caveat we cannot fix from our side: whoever operates the assistant you are using can see your conversation with it. That is between you and them. Our half of the exchange keeps no record.

If a search comes back busy

Keyword search over half a million opinions is the one expensive thing here, so a small number of searches run at once and the rest are asked to retry rather than queued. If you see that, narrow the query — a broad common-law term matches tens of thousands of cases and is rarely what you wanted anyway. The other five tools are unaffected and answer in milliseconds.

Bulk access

If you want the corpus rather than a lookup tool, don't scrape it — ask us and we will send you a copy. There is a Minnesota bundle on the Internet Archive already, and other cuts are free for the asking. It is public-record data; a better door is a more useful answer than a lock.

— DocketDrift