MCP 服务器

Epstein-Files-Plugin

Enables AI agents to search and query the public DOJ Epstein Files release across 20 databases, resolve EFTA numbers to official DOJ PDFs, and cross-check viral claims against a fact-checked registry.

README

<br />

<b>Epstein Files Plugin</b> — query the DOJ Epstein Files Transparency Act release from any AI agent

<p align="center"> <a href="#quickstart">Quickstart</a> · <a href="#architecture">Architecture</a> · <a href="#the-11-read-only-mcp-tools">Tools</a> · <a href="#use-with-ai-agents">Use with AI Agents</a> · <a href="#examples">Examples</a> · <a href="#responsible-use">Responsible Use</a> </p>

Turns a name, date, keyword, EFTA page number, or an open-ended question — about the email corpus, the photos, the transcripts, or the knowledge graph — into the exact, citable, official DOJ document. No hand-paging through ~2.8 million pages. No fighting DOJ's bot-blocked search.

</div>

Operates only on already-public DOJ court records and a public full-text mirror. Built for research, fact-checking, and source-location — not re-identification, harassment, or doxxing. See Responsible Use.

About

Epstein Files Plugin indexes and resolves documents in the public DOJ Epstein Files Transparency Act release (Public Law 119-38, signed 2025-11-19) — roughly 1.4M documents / 2.8–2.9M pages published across 12 DataSets at justice.gov/epstein, mirrored and indexed as 20 queryable databases by the third-party service epstein-data.com (both figures verified live 2026-08-24; DOJ's total collected corpus is larger still, ~6M pages, of which this is the portion published so far — re-verify before citing as exhaustive).

Available as:

🧩 a Claude Code plugin — installed from a marketplace in two lines
🔌 an MCP server (.mcp.json) — the same server, standalone — local stdio, no API keys, no remote host, works with any MCP-compatible client: Claude Desktop, Cursor, Windsurf, or a custom agent runtime built on the official mcp SDK
💻 a Python CLI — python src/efta_researcher.py, no MCP client required
📦 a Python library — from efta_core import efta_to_url, get_dataset, import directly

Why use it:

🚀 Faster than manual lookup. DOJ's own search is blocked for automated/headless access (Akamai bot protection — see Responsible Use); this goes straight from a name, date, or keyword to the matching EFTA number, in seconds.
🎯 Direct-to-source, every time. Every result resolves to the official justice.gov/epstein PDF URL — you verify the primary document yourself, never a paraphrase of it.
✅ Fact-checked, not sensationalized. The known-document registry ships with sourced fact-checks (EUvsDisinfo, Reuters, Tempo, and others) so a "connection" is labeled for what the document actually shows, not for what a viral caption claims.
🧠 Completely queryable, not just the inbox. Two purpose-built tools cover the email corpus; five generic tools reach all 20 databases — images, transcripts, OCR text, handwriting, depositions, the knowledge graph — plus a read-only SQL escape hatch for anything the filter API can't express.
🗺️ Structure, reconstructed once. The 12-DataSet / EFTA-Bates-number boundary table was cross-checked against DOJ's own disclosure pages, so you don't re-derive it per query.
🔓 No lock-in. The same lookups work from a CLI, a Python import, or any MCP client.

Architecture

flowchart LR
    subgraph Client["Any MCP client"]
        A["Claude Code / Claude Desktop\nCursor · Windsurf · custom agent"]
    end

    subgraph Server["epstein-files-plugin — local stdio, no auth"]
        direction TB
        PB["Purpose-built tools\nefta_search · efta_filter_email\nefta_lookup · efta_known_docs"]
        GEN["Generic tools\nefta_list_databases · efta_list_tables\nefta_describe_table · efta_query_table\nefta_run_sql"]
        LOCAL["Local-only\nefta_get_url — pure computation,\nno network call"]
    end

    subgraph Upstream["Third-party mirror"]
        DS[("epstein-data.com\nDatasette JSON API\n20 databases")]
    end

    DOJ[("justice.gov/epstein\nofficial PDF source")]

    A -- "MCP / stdio\nJSON-RPC" --> Server
    PB --> DS
    GEN --> DS
    LOCAL -.->|"constructs URL,\nno request sent"| DOJ
    DS -.->|"indexes"| DOJ

    style Server fill:#1a2332,stroke:#4b8bbe,color:#e8edf4
    style Upstream fill:#0f1720,stroke:#2dd4a7,color:#e8edf4
    style DOJ fill:#0f1720,stroke:#94a3b8,color:#e8edf4

Two tool layers, deliberately: purpose-built wrappers over the single most common table (fast, ergonomic, proper pagination), and a generic introspection/query/SQL layer that reaches every other database without a dedicated tool per table — the answer to "can everything be queried," and what keeps this working as epstein-data.com's schema evolves through 2031 without new code. See CLAUDE.md for the full design rationale.

I want to...

Goal	Use this
Find a document by name, date, or keyword	`efta_search` / `efta_filter_email`
Turn a citation (`EFTA00741068`) into an official DOJ link	`efta_get_url`
Check whether a viral claim is real	`efta_known_docs` → Research & Fact-Check Posture
Search images, transcripts, OCR text, or the knowledge graph	`efta_list_databases` → `efta_query_table`
Run a join, a count, or a `GROUP BY`	`efta_run_sql`
Wire this into an agent that isn't Claude Code	Use with AI Agents
Understand the EFTA numbering / DataSet system	DataSet Reference

Quickstart

Prerequisites

Python 3.11 or later (3.10 reaches end-of-life 2026-10-31 — see Tech stack currency)
pip

Install as a Claude Code plugin

/plugin marketplace add Zereo0317/Epstein-Files-Plugin
/plugin install epstein-files-plugin@epstein-files-plugin

Install for CLI / standalone MCP / library use

git clone https://github.com/Zereo0317/Epstein-Files-Plugin.git
cd Epstein-Files-Plugin
python -m pip install --upgrade pip
pip install -r requirements.txt   # requests, mcp (pinned <2.0 — see below)

Run the CLI

python src/efta_researcher.py --list
python src/efta_researcher.py --search "trilateral commission"
python src/efta_researcher.py --sender epstein --recipient schank --date 2009-10-23
python src/efta_researcher.py --efta EFTA00741068
python src/efta_researcher.py --databases
python src/efta_researcher.py --tables image_analysis
python src/efta_researcher.py --sql "select dataset, count(*) as n from doc_search group by dataset" --database full_text_corpus

Run the MCP server standalone

python src/mcp_server.py

The 11 read-only MCP tools

🔎 Purpose-built (the common case — full-text + email metadata):

Tool	Purpose
`efta_search(query, limit, cursor)`	Substring search across the Datasette index — true total match count + `cursor` paging
`efta_filter_email(sender, recipient, date_exact, date_prefix, subject, limit, cursor)`	Filter emails by metadata fields — same total-count + paging
`efta_known_docs(category)`	List pre-verified, fact-checked known documents
`efta_get_url(efta_number)`	Convert an EFTA number to its official DOJ PDF URL
`efta_verify_url(efta_number)`	HEAD-check whether a DOJ PDF URL is live (see caveat below)
`efta_lookup(efta_number)`	Full document metadata from the Datasette index

🧬 Generic (every one of the 20 databases, no hardcoding per table):

Tool	Purpose
`efta_list_databases()`	List all 20 Datasette databases (images, transcripts, OCR, depositions, ...)
`efta_list_tables(database)`	List every table in one database, with columns + row counts
`efta_describe_table(database, table)`	Column list + row count for one table
`efta_query_table(database, table, filters, limit, cursor)`	Filter-suffix query against any table
`efta_run_sql(database, sql, params, limit)`	Read-only SQL — joins, aggregation, `GROUP BY`; Datasette rejects any non-`SELECT` with HTTP 400

✅ All 11 tools are annotated readOnlyHint/idempotentHint (and openWorldHint on everything network-facing) per the MCP tool-annotations convention — a client can safely auto-run them.

⚠️ Pagination: limit caps at 50–100 depending on the tool. A query can match far more (e.g. "pizza" currently matches 233 documents) — paged tools surface (N of TOTAL shown) and, when more exist, a cursor to continue.

⚠️ URL verification: justice.gov gates every PDF behind an age-verify + Akamai challenge, so a raw HTTP status can't reliably distinguish live from missing. URL correctness comes from the verified DataSet boundary table below, not from probing justice.gov.

❌ Not covered by the purpose-built tools alone: image_analysis (92K captioned images), transcripts (435 audio/video transcripts), knowledge_graph, ocr_database, handwriting_transcriptions, and 13 more — all reachable via the generic tools instead. Call efta_list_databases() for the live, current list rather than trusting this table.

Use with AI Agents

Epstein Files Plugin is MCP-first, not Claude-first: a standard local stdio server (built on the FastMCP class bundled inside the official mcp Python SDK) speaking plain Model Context Protocol. It works with any MCP-compatible client.

Client	How it connects
Claude Code / Claude Desktop	`/plugin marketplace add` + `/plugin install`, or the raw `.mcp.json`
Cursor / Windsurf / Cline	Add the server entry from `.mcp.json` to the client's MCP settings
OpenClaw	Add the same generic config below under `mcpServers` in your own `~/.openclaw/openclaw.json` (or `openclaw config set mcpServers.epstein-files-plugin.command "python"` etc.) — a local stdio server needs no `transport` field, OpenClaw auto-detects it from `command`. This plugin's own ClawHub listing manifest can't auto-wire this for you (OpenClaw's plugin manifest has no MCP-server field as of the current release — see `CLAUDE.md`); this per-user config is the real, working path.
ChatGPT, Gemini, or any custom/headless agent	Point it at `src/mcp_server.py` over stdio — no plugin system or Claude dependency required

Generic MCP client configuration

{
  "mcpServers": {
    "epstein-files-plugin": {
      "command": "python",
      "args": ["/absolute/path/to/epstein-files-plugin/src/mcp_server.py"]
    }
  }
}

No API keys, no auth, no remote server — a local process talking stdio, identical behavior in every client.

Claude Code

Installed as a plugin, the same server auto-loads from this repo's .mcp.json (uses ${CLAUDE_PLUGIN_ROOT}, no path editing needed):

{
  "mcpServers": {
    "epstein-files-plugin": {
      "command": "python",
      "args": ["${CLAUDE_PLUGIN_ROOT}/src/mcp_server.py"],
      "env": { "PYTHONUNBUFFERED": "1" }
    }
  }
}

Environment overrides (resilience against a mirror or domain change)

Variable	Default	Affects
`EFTA_DATASETTE_BASE_URL`	`https://epstein-data.com`	Every search/query/SQL tool
`EFTA_DOJ_BASE_URL`	`https://www.justice.gov`	`efta_get_url`, `efta_verify_url`, the CLI's `--download`/`--verify`

A future change to either third-party host is a config change, not a code change.

Examples

Example 1 — Resolve a citation to its official source

User request:  "What's the DOJ URL for EFTA00741068, and which DataSet is it in?"

Response:
  EFTA: EFTA00741068
  DataSet: DS9
  URL: https://www.justice.gov/epstein/files/DataSet%209/EFTA00741068.pdf

Under the hood:
  efta_get_url("EFTA00741068") parses the Bates number, resolves it against the verified
  12-DataSet boundary table, and constructs the official DOJ URL — no network call, no guessing.

Example 2 — Find an email by sender, recipient, and date

User request:  "Find the Epstein -> Roger Schank email from October 23, 2009."

Response:
  Found 3 email(s):
    EFTA00741068  DS9   2009-10-23  12:01:12   <- primary
    EFTA00885615  DS9   2009-10-23  12:01:12   (OCR duplicate, reads "grnail.com")
    EFTA01821140 DS10   2009-10-23  12:01:12   (third copy, later processing batch)

Under the hood:
  efta_filter_email(sender="epstein", recipient="schank", date_exact="2009-10-23") queries the
  epstein-data.com Datasette API (DOJ's own /multimedia-search is blocked for headless clients)
  and returns every metadata match, so duplicate copies can be cross-referenced by timestamp.

Example 3 — Search beyond the email corpus (generic tools)

User request:  "Any photos in the release that show a passport?"

Response:
  efta_list_tables("image_analysis") -> "images" table, 92,249 rows, column "analysis_text"
  efta_query_table("image_analysis", "images", filters={"analysis_text__contains": "passport"})
  -> matching rows with efta_number, source_pdf, and the analysis text itself

Under the hood:
  image_analysis isn't reachable through efta_search (that only covers full_text_corpus).
  efta_query_table works against any of the 20 databases using the same filter-suffix syntax,
  so no dedicated "image search" tool was needed.

Example 4 — A question the filter API can't express (raw SQL)

User request:  "Break down the document count by DataSet."

Response:
  efta_run_sql("full_text_corpus",
    "select dataset, count(*) as n from doc_search group by dataset order by dataset")
  -> 1:650, 2:150, 3:57, 4:143, 5:82, 6:13, 7:17, 8:10479, 9:480658, 10:496404,
     11:331597, 12:12339, 98:6, 99:23210   (live counts, 2026-08-24)

Note: two values (98, 99) fall outside the documented 1-12 DataSet scheme — small catch-all
buckets in the source data itself. efta_get_url()/get_dataset() only resolve DataSets 1-12.

DataSet Reference

EFTA numbers are page (Bates) identifiers, not document identifiers — a 20-page PDF consumes 20 consecutive EFTA numbers. Boundaries below are the forensic per-file ranges from the rhowardstone/Epstein-research-data mapping, cross-checked against DOJ's own disclosure pages.

DataSet	EFTA range	Contents
DS01	1 – 3,158	Photos, physical scans
DS02	3,159 – 3,857	Photos, seized scans
DS03	3,858 – 5,586	Grand jury exhibits
DS04	5,705 – 8,320	Records, court filings
DS05	8,409 – 8,528	Seized scans, depositions
DS06	8,529 – 8,998	Depositions, indictments
DS07	9,016 – 9,664	Transcripts
DS08	9,676 – 39,023	Emails, police reports
DS09	39,025 – 1,262,781	Main email corpus
DS10	1,262,782 – 2,205,654	Emails, financial
DS11	2,205,655 – 2,730,264	Emails, device data
DS12	2,730,265 – 2,858,497	Court filings, FBI + expansion

DOJ URL pattern: https://www.justice.gov/epstein/files/DataSet%20{N}/EFTA{efta:08d}.pdf

Research & Fact-Check Posture

This release attracts conspiracy framings. Every entry in the known-document registry (efta_known_docs) is confidence-tagged and de-sensationalized:

Topic	What the documents show
Trilateral Commission / CFR	Epstein's own bio listed him as a former member — elite networking, not a plot
Rothschild	A real advisory relationship (~$25M Southern Trust agreement); the "Ukraine upheaval" email — EUvsDisinfo flagged the "coup" version as disinformation
Rockefeller	A Rockefeller University board seat + donor relationship — institutional, not "bloodline"
Illuminati	An inbound, unsolicited email sent to Epstein; no reply on record; not evidence of membership
Gates / BGC3	A real pandemic-preparedness scope document; fact-checkers found no COVID-19 planning link

Responsible Use

✅ Operates only on already-public DOJ releases at justice.gov/epstein and a public third-party full-text mirror (epstein-data.com). No private data, no paywalled sources, no scraping behind a login.
✅ Every claim in the known-document registry carries a source and a confidence tag — a "connection" is labeled for what a document literally shows, never for what a viral caption claims.
✅ efta_run_sql broadens what can be queried, not the ethical posture: Datasette's own API only accepts SELECT (a non-SELECT is rejected with HTTP 400 before it reaches SQLite — verified live), and it reaches no data epstein-data.com doesn't already expose to anyone browsing its site directly.
❌ Not for re-identification, harassment, or doxxing. This is source-location and verification tooling, not an investigation or accusation engine — it does not allege wrongdoing beyond what a document shows.
❌ Not legal advice, not an official DOJ product, and not affiliated with epstein-data.com — an independent client of their public API.

Tech stack currency (2026-08-24)

Python: 3.11+ required (bumped from 3.10 — EOL 2026-10-31). Tested against 3.14.7.
MCP SDK: pinned mcp>=1.29.0,<2.0.0. The official SDK's v2.0.0 (2026-07-28) renamed mcp.server.fastmcp.FastMCP to mcp.server.mcpserver.MCPServer — a breaking change this server hasn't migrated to. An unpinned mcp>=1.0.0 would silently resolve to v2.x and fail to import. The unrelated standalone fastmcp PyPI package (PrefectHQ, now v3.x/4.0) is not a dependency.
pip: install command above runs python -m pip install --upgrade pip first.
requests: >=2.31.0, no known constraint against newer 2.x releases.

Project Layout

.claude-plugin/plugin.json   Plugin manifest (Claude Code convenience only)
.mcp.json                    Local stdio MCP server config (client-agnostic)
src/efta_core.py             DataSet boundary table, EFTA -> URL, known-document registry
src/epstein_datasette.py     epstein-data.com Datasette API client (purpose-built + generic layer)
src/doj_auth.py              justice.gov public anti-bot challenge helper + verification
src/efta_researcher.py       Standalone CLI
src/mcp_server.py            FastMCP server exposing the 11 tools above
skills/efta-research/        Claude Code skill: research methodology (optional convenience)
skills/doj-auth/             Claude Code skill: justice.gov access details (optional convenience)

Community & Support

Contributing

This is a public, single-maintainer research tool (Zereo0317/Epstein-Files-Plugin). Issues and pull requests are welcome.

License

MIT-0 (MIT No Attribution) — see LICENSE. Chosen over plain MIT specifically for ClawHub compatibility, which requires MIT-0 with no per-skill overrides. The repository is public on GitHub; the license grant governs redistribution/reuse of the code by anyone.

Disclaimer

Epstein Files Plugin only surfaces documents the DOJ has already made public under the Epstein Files Transparency Act, plus a public third-party full-text mirror (epstein-data.com) of that same release. It resolves citations to their official source and reports what a document literally contains — it does not conduct original investigation, does not allege wrongdoing beyond what a document shows, and does not host, re-host, or expose any non-public data. Intended for research, fact-checking, and source verification — not re-identification, harassment, or doxxing.