Grateful Dead Knowledge Graph
675,641 tracks. 18,106 recordings. 2,336 shows.
30 years of the Grateful Dead, structured and searchable.
A knowledge graph built from the Internet Archive's Grateful Dead collection, served through an MCP server that Claude Desktop can query with natural language.
What Is This
DeadGraph is a structured knowledge base of the Grateful Dead's entire live concert history. Every show from 1965 to 1995. Every recording uploaded by the community to the Internet Archive. Every track on every tape, normalized into canonical entities and made queryable.
The data is harvested from the Internet Archive's Grateful Dead collection - tens of thousands of community-contributed recordings that are freely accessible. DeadGraph doesn't redistribute audio. It structures the metadata: who played where, when, what songs, in what order, for how long.
The whole thing is served through an MCP (Model Context Protocol) server. Connect it to Claude Desktop and ask questions in plain English. No SQL required. No special syntax. Just ask like you're talking to a fellow Deadhead.
2,336 unique shows spanning 1965-1995. 3,144 raw venue strings collapsed into canonical show records via text clustering.
18,106 recordings catalogued with source type, transfer quality, community ratings, and direct links back to the Internet Archive.
Every song on every tape. With durations, segue detection, and statistics: first play, last play, performance count, average length.
Ask Claude Desktop anything about the Dead's live history. DeadGraph handles the lookup via five purpose-built MCP tools.
The Pipeline
Six stages transform raw Internet Archive metadata into a fully indexed, rights-aware, edge-deployed knowledge base.
Async scraping of 18,106 items from the Internet Archive API using httpx with tenacity retry logic (5 attempts, exponential backoff). Pulls metadata, recording sources, track listings, and community ratings. Outputs two raw JSONL files.
Python / httpx / asyncio
3,144 raw venue strings collapsed into 2,336 unique shows via slug-based clustering
and date grouping. Each show gets a stable ID in the format
gd:show:1977-05-08:cornell-university. Duration parsing normalizes
every time format the Archive uses.
Song-level statistics computed across all 675,641 tracks: first known performance, last known performance, total count, average duration. Segue detection identifies song transitions using track-order analysis. All stats stored in a single enriched stats.json.
Python / statisticsRights-aware filtering via policy.yaml. Only recordings with CC0, CC-BY, CC-BY-NC, or Public Domain licenses pass through. Blocked sources are excluded. No audio redistribution. No lyrics. Metadata as facts with citations - that's the policy.
Python / policy.yamlFiltered JSONL loaded into Cloudflare D1 (SQLite at the edge) with FTS5 full-text search indexes on show fields, venue names, song titles, and recording identifiers. Schema migrations run via Wrangler. Zero cold-start overhead.
Cloudflare D1 / SQLite FTS5 / WranglerCloudflare Worker with Hono serving the MCP protocol over HTTP with Server-Sent Events. OAuth 2.1 authentication. Five MCP tools for natural language queries. Global edge deployment on Cloudflare's network. $0/month on the free tier.
TypeScript / Hono / Cloudflare Workers / OAuth 2.1Five MCP Tools
Connect DeadGraph to Claude Desktop and query 30 years of concert history in plain English. Five purpose-built tools handle every kind of question.
Tool 01
search_showsSearch shows by venue, date range, or keywords using FTS5 full-text search. Returns shows with recording counts, sorted by relevance.
Example queries
"cornell 1977"
"winterland new year"
"1969-02"
Tool 02
get_showFull details for a specific show - venue, date, musicians listed, all recordings sorted by community rating. Each recording links directly to Archive.org.
Example
get_show("gd:show:1977-05-08:cornell-university")
Tool 03
get_setlistFull track listing for a specific recording. Song titles, individual durations, total runtime, and a direct link to stream the recording on Archive.org.
Example
get_setlist("gd:rec:gd1977-05-08.sbd")
Tool 04
search_songsFind every known performance of a song across all 30 years. Returns unique show dates, performance counts, first and last known performance, and average duration.
Example
"Dark Star" - 233 show dates found
"Scarlet Begonias"
Tool 05
get_statsDatabase summary statistics at a glance. Total shows, recordings, tracks, date range covered, and a breakdown of recording source types (SBD, AUD, MTX).
Returns
2,336 shows / 18,106 recordings
675,641 tracks / 1965-1995
In Action
These placeholders will be replaced with actual Claude Desktop screenshots. For now - here's what you can ask.
What shows did the Dead play at Cornell?
Screenshot coming soon
Find the longest Dark Star performances
Screenshot coming soon
What was the setlist for 5/8/77?
Screenshot coming soon
How many times did they play Scarlet Begonias into Fire on the Mountain?
Screenshot coming soon
The Numbers
Tracks
songs across all recordings
Recordings
community-uploaded tapes
Shows
unique concert dates
Years Active
30 years of live music
Tests
100% pass rate
Monthly Cost
Cloudflare free tier
Standing on Shoulders
In 1987, three Deadheads - John W. Scott, Stu Nixon, and Mike Dolgushkin - published DeadBase, the first comprehensive catalog of every Grateful Dead show ever played. The band gave its blessing. Vault archivist Dick Latvala contributed directly. Proceeds went to the Rex Foundation.
For nearly two decades, DeadBase was how Deadheads navigated the tape trading world. You looked up a song, found every date it was played, then sought out the recording. Nineteen editions were published across 28 years, culminating in DeadBase 50 (2015) - a 992-page book covering through the Fare Thee Well shows. Data was collected from hundreds of fans via Usenet, the WELL, and mailed-in corrections. It was the community's shared factual canon before the internet made that easy.
DeadBase documented 2,318 shows and 480 distinct songs across 36,504 catalogued performances. It was a setlist reference - definitive, meticulous, and limited by its medium. The original database was eventually lost; the final edition is scanned PDFs, not live data.
DeadGraph picks up where DeadBase left off. Same obsessive cataloging instinct, different era. Where DeadBase tracked setlists, DeadGraph tracks recordings - 18,106 of them, with source type, lineage, community ratings, and 675,641 individual tracks. Where DeadBase required page-flipping, DeadGraph answers natural language questions. Where DeadBase was frozen in print, DeadGraph is a live pipeline that can re-harvest and re-index.
| DeadBase X (1997) | DeadGraph (2026) | |
|---|---|---|
| Shows | 2,318 | 2,336 |
| Song performances | 36,504 | 675,641 tracks |
| Recordings | Notes only | 18,106 with metadata |
| Search | Page flip | FTS5 + natural language |
| Updates | New print run | Pipeline re-run |
| Format | Print / PDF | Live API |
| Cost | $90 | Free |
DeadBase built the foundation. The Internet Archive preserved the tapes. DeadGraph connects them to the age of AI. Different tools, same community, same obsession.
Under the Hood
Python for data wrangling. TypeScript for the edge. Cloudflare for everything in between.
Python Pipeline
httpx, tenacity, FAISS, MiniLM - async harvest to indexed knowledge base
Cloudflare Workers
TypeScript MCP server with Hono framework - edge compute, zero cold start
Cloudflare D1
SQLite at the edge with FTS5 full-text search indexes
OAuth 2.1
Standard auth for MCP connections from Claude Desktop and other clients
Internet Archive
Data source - 18,106 freely accessible community recordings
MCP Protocol
Model Context Protocol - standard interface for AI tool integrations
Get Started
Three steps to start querying 30 years of Grateful Dead history from Claude Desktop.
Open the Claude Desktop app. Click the hamburger menu (top left), then
Settings and Developer. Click
Edit Config to open claude_desktop_config.json.
Paste this configuration into your mcpServers block:
{
"mcpServers": {
"deadgraph": {
"command": "npx",
"args": [
"mcp-remote",
"https://mcp.deadgraph.com/sse"
]
}
}
}Restart Claude Desktop. On first connection, a browser window opens for OAuth authentication. Enter the access PIN when prompted, and you're connected. Start asking questions.