DDX for PostgreSQL — About
WHAT IS THIS?
The Database Development neXus: for PostgreSQL (DDX for PostgreSQL, served at pg.ddx.io) is a free, read-only archive and intelligence service for the PostgreSQL community.
All data served is already public. We add full-text search, semantic search, code intelligence with
call graphs (get_callers and get_callees), and structured MCP access for AI agents.
No personal data is collected. No accounts exist. All protocols are read-only. The software powering this service (the open-source pg.ddx.io server codebase) is a derivative of public-inbox.org, reimplemented and significantly extended.
WHY THIS EXISTS
PostgreSQL's mailing lists, git repositories, documentation, wikis, and commitfest patches contain decades of distributed knowledge. We currently index over 1.8 million messages spanning 1985-present across 85 community archives (pgsql-hackers alone holds about 650,000), with historical backfill ongoing. This knowledge is scattered across different systems and formats, making it expensive and time-consuming for AI systems (and humans) to:
- Search across all sources simultaneously
- Correlate information (which discussion led to this commit?)
- Fetch and parse different formats (mbox, git, HTML, markdown)
- Handle repeated requests that overload upstream infrastructure
DDX for PostgreSQL solves this by:
- Indexing all PostgreSQL community sources in one place
- Providing full-text search, fuzzy matching, and semantic search, plus a natural-language Ask the Archive feature that answers questions with citations to the source messages
- Offering multiple access protocols (HTTP, NNTP, IMAP, POP3, MCP, Git)
- Reducing load on postgresql.org infrastructure
FOR AI/LLM DEVELOPERS
The Problem
Fetching raw PostgreSQL data inefficiently wastes LLM tokens. Example:
- Before: Fetch all 100,000 pgsql-hackers messages, parse manually → 50MB+ of tokens
- After: Query structured, indexed results → 2MB of tokens (25x cheaper)
DDX for PostgreSQL provides pre-indexed, pre-embedded data with structured queries, so you get exactly what you need without parsing overhead.
How well the search tools find the right thread, measured against public answers, is on /status.
Connect your agent
The endpoint is https://pg.ddx.io/mcp. No key, no account.
Claude Code
claude mcp add --transport http pg-ddx https://pg.ddx.io/mcp
Codex
codex mcp add pg-ddx --url https://pg.ddx.io/mcp
Or in ~/.codex/config.toml:
[mcp_servers.pg-ddx] url = "https://pg.ddx.io/mcp"
GitHub Copilot CLI
copilot mcp add --transport http pg-ddx https://pg.ddx.io/mcp
This writes ~/.copilot/mcp-config.json:
{"mcpServers": {"pg-ddx": {"type": "http",
"url": "https://pg.ddx.io/mcp", "tools": ["*"]}}}
VS Code
In .vscode/mcp.json (VS Code uses servers, not mcpServers):
{"servers": {"pg-ddx": {"type": "http", "url": "https://pg.ddx.io/mcp"}}}
Cursor
In ~/.cursor/mcp.json:
{"mcpServers": {"pg-ddx": {"url": "https://pg.ddx.io/mcp"}}}
Zed
In settings.json:
{"context_servers": {"pg-ddx": {"url": "https://pg.ddx.io/mcp"}}}
OpenCode
In opencode.json:
{"$schema": "https://opencode.ai/config.json",
"mcp": {"pg-ddx": {"type": "remote", "url": "https://pg.ddx.io/mcp", "enabled": true}}}
Any other client
MCP Streamable HTTP at https://pg.ddx.io/mcp; clients that only speak the
legacy SSE transport use https://pg.ddx.io/mcp/sse. No key, no account.
/mcp lists the core tools; every tool (129) is at /mcp/full.
Reference: /mcp-tools.
How Your Agent Uses This
You'll never need to do this yourself! This is just a peek under the covers. Your AI agent will use the methods exported by the MCP service to invoke this service and get information about Postgres faster than you'd imagine. Just give it a try!
When you connect your agent to DDX for PostgreSQL via MCP, it gains access to 100+ structured methods for searching mailing lists, analyzing code, exploring git history, and discovering commitfest patches. The agent calls these directly—no manual HTTP requests needed.
Behind The Scenes (For Curiosity)
If you want to see how it works, here are some example queries against the live HTTP/JSON surface. The agent never makes you type these — it calls equivalent MCP tools directly.
1. Search a mailing list and get JSON results:
curl -s 'https://pg.ddx.io/m/pgsql-announce/?q=pgbouncer&format=json' | jq .
Returns matching messages with subject, from, date, message-id, and a thread cursor. Each indexed inbox has its own search endpoint at /m/<inbox>/?q=<term>&format=json (the canonical path; /<inbox>/?... still 301-redirects there). Drop &format=json to get the HTML view a browser would render.
2. Fetch a single message as JSON:
curl -s 'https://pg.ddx.io/m/pgsql-announce/177913516055.803.10890499456720421868@wrigleys.postgresql.org/info.json' | jq .
Returns the message's headers, decoded body and attachments; follow up with its thread (/t/), raw bytes (/raw) or the list's Atom feed.
3. Talk to the MCP server directly (JSON-RPC over HTTP):
curl -s -X POST -H 'Content-Type: application/json' \
-d '{"jsonrpc":"2.0","id":0,"method":"initialize",
"params":{"protocolVersion":"2025-11-25",
"capabilities":{},
"clientInfo":{"name":"curl","version":"1"}}}' \
-H 'Accept: application/json, text/event-stream' https://pg.ddx.io/mcp | jq .
Initializes a session and returns the server's protocol version, capabilities, and tool surface. Real agent clients (Claude Desktop, Cursor, custom integrations) do this transparently. Subsequent tool calls (tools/list, tools/call, resources/read) require the session ID returned here, which is why most users let an MCP client library handle the protocol.
4. Clone an inbox via Git smart-HTTP:
git clone https://pg.ddx.io/m/pgsql-announce.git
Each inbox is also a public-inbox v2 git repository — mirror the whole archive locally and run your own indexer on top.
Why This Matters
- Context Windows Are Expensive: Structured queries cost 25-100x less in tokens than raw data
- Semantic Search & Ask the Archive: Ask "How does PostgreSQL handle NUMA?" and find related code, discussions, and documentation automatically. The Ask the Archive page answers plain-English questions grounded only in the mailing-list archive, with citations to the source messages (hybrid semantic + keyword retrieval, GPU-synthesized).
- Real-Time Data: Updates hourly (mailing lists), continuously (git), nightly (docs)
- No Overhead: Data is pre-parsed, pre-embedded, ready to feed to your LLM
INFRASTRUCTURE
Open source. Self-hosted on one dedicated Hetzner server (hel1: i7-7700, 4 cores / 8 threads, 64 GB RAM, GTX 1080 GPU, Helsinki), the only host since the Exoscale hosts were closed on 2026-07-15; there is no failover host. All ingestion, indexing, and vector embedding is performed on-premise — no third-party service sees the content of any query, nor any private data about who asked it. The full source for the application and its deployment is at codeberg.org/ddx.
LIVE ACTIVITY (IRC)
The site posts deploy alerts, status changes, and workflow events to a public, read-only IRC channel. If you'd like to lurk on what's going on behind the scenes, connect with any IRC client.
| Server | irc.ddx.io (TLS only on tcp/6697) |
| Channel | #ddx |
| Voiced (writes) | pgesq-bot (alert poster) and gregburd (operator). Everyone else can join and read; the channel is +m. |
| Topic | DDX for PostgreSQL — site alerts, status, workflow. |
Example invocations:
irssi:
/network add -nick lurker ddx /server add -auto -ssl -ssl_verify -network ddx irc.ddx.io 6697 /channel add -auto #ddx ddx /connect ddx
weechat:
/server add ddx irc.ddx.io/6697 -tls /set irc.server.ddx.autojoin "#ddx" /connect ddx
Browser (no install, lurker mode):
https://kiwiirc.com/nextclient/?server=irc.ddx.io:+6697&channel=%23ddx&nick=lurker_?
Anyone can join and listen. Only allowlisted SASL-authenticated nicks can send messages.
DATA SOURCES
- Mailing lists: 85 PostgreSQL community archives, updated hourly, and the PostGIS lists (postgis-devel, postgis-users from lists.osgeo.org), updated daily. Full NNTP/IMAP/POP3/Git/HTTP access.
- GitHub issues and pull requests: projects whose record is their GitHub tracker rather than a mailing list, each an inbox of mail threads (one per issue or PR, with its comments, reviews and inline review comments), synced hourly: pgjdbc, psqlodbc, aws-advanced-jdbc-wrapper, psycopg, pgbouncer, pgdog, pgbackrest, patroni, pg_repack, pgaudit, pg_partman and pgvector.
- Git repositories: PostgreSQL upstream, the commitfest bot's patch branches (their heads), connection poolers (pgbouncer, pgpool2, pgdog), drivers (JDBC, the AWS JDBC wrapper, ODBC, psycopg), and the tools and extensions above plus PostGIS (pgbackrest, Patroni, pg_repack, pgAudit, pg_partman, pgvector).
- Documentation: postgresql.org/docs (all versions), full-text indexed
- Wiki: wiki.postgresql.org (1,800+ public pages), searchable
- Commitfest: Patch submission lifecycle tracking, in development — data is indexed and queryable via MCP; the /commitfest web UI is not yet live
- Buildfarm: 257 animals + run history, indexed and queryable via MCP and
/mcp/api/buildfarm/* - Discord: PostgreSQL community Discord — indexed for cross-source identity linkage only. Participants are connected to their git/email/docs activity via their public handle; per our agreement with the Discord community, message content is never exposed through any API.
The critical value is the connective tissue between these sources: linking a mailing list discussion to the commit it produced, to the code symbols it changed, to the documentation it updated, and to the build results that validated it.
LEGACY: THE BERKELEY POSTGRES ARCHIVE
The /legacy section is a static mirror of the original Berkeley POSTGRES archive at dsf.berkeley.edu/postgres.html — Stonebraker's research project from 1986–1995 that became modern PostgreSQL. We mirror it here for permanence; bytes are preserved in our R2 backup as well as served from DDX for PostgreSQL.
The mailing-list discussions, source releases (POSTGRES v3 through v4.2
plus postgres95-{0.01..1.02}), patches, and papers are also indexed and
queryable through MCP and search_docs. See /legacy.
POSTGRES was Berkeley's successor to University INGRES (1973–1988).
/legacy/ingres mirrors the INGRES archive at
s2k-ftp.cs.berkeley.edu/ingres: the 8.9
release, its 1993 Linux port, patches, the reference manual and "The Design and
Implementation of INGRES" (1976). The manuals are searchable as documentation version
ucbingres, and the code as the repositories ucb-ingres-8.9 and
ucb-ingres-8.9-linux.
EXTENSIONS & ACKNOWLEDGMENTS
DDX for PostgreSQL is a thin service (built on the open-source pg.ddx.io server codebase) over a much larger pile of work done by other people. The PostgreSQL extensions listed below do most of the actual heavy lifting; the service just routes traffic and stitches results.
| Extension | Version | Role |
|---|---|---|
| PostgreSQL | 18.6 | Core database — every byte of the archive lives in PostgreSQL. |
| pg_turbovec | 2.15.0 | TurboQuant-quantised vector index access method (pgrx). Backs all four live embedding columns (ag_messages, ag_embeddings, ag_ci_embeddings, ag_document_embeddings) with 4-bit-quantised ANN. Drives semantic search over messages and code symbols. The v1.27 on-disk format de-duplicates the quantised codes, roughly halving index size at identical recall. |
| pg_fts | 1.16.0 | BM25 full-text relevance ranking with a dedicated inverted-index access method (fts). Drives search_messages and search_symbols. Crash-safe and streaming-replica-safe via GenericXLog; supports boolean / phrase / prefix / fuzzy / regex queries. |
| pg_tre | 4.6.1 | TRE-pattern regex search with edit-distance (drives ?q=…®ex=1&k=N). Built on Ville Laurikari's TRE library. |
| pg_mentat | 1.19.1 | EDN/datalog triple store; lets us cross-join mailing lists, git, docs, and wiki in one query language. |
| pg_cron | 1.6.8 | In-database job scheduler — mailing-list ingest, git pulls, backups all run from here. |
| pgcrypto | 1.4 | Hashing for content fingerprints (deduping messages, naming git objects). |
| pg_stat_statements | 1.12 | Query telemetry; how we find slow queries before users do. |
| PgQue † | 0.2.2 | Snapshot/batch job queue (pure plpgsql, no C extension). pg_cron triggers job_management.enqueue_job() which writes to pgque.event_*; agora-job-dispatcher.service consumes batches and fires the corresponding systemd *.service. Replaces the prior in-DB job_management.job_queue table-as-queue. Apache-2.0. |
† PgQue is a pure plpgsql schema add-on, not a binary PostgreSQL extension — it doesn't appear in \dx. Listed here because it carries weight equivalent to one.
Also indispensable, even though they're not loaded as PostgreSQL extensions:
- public-inbox v2 — the on-disk archive format. The pg.ddx.io server reimplements public-inbox semantics; we owe Eric Wong and the public-inbox contributors the entire data model.
- The pg_fts author for the BM25 full-text search extension.
- Cloudflare Pages / Workers / Load Balancer — the edge proxy in front of pg.ddx.io.
- ollama — local embedding inference; no third-party model provider sees query text.
- Linux / NixOS / colmena — the deployment substrate. The whole stack is one flake colmena pushes to the host.
- DB-IP Lite IP-to-Country database — the IP-to-country mapping that powers the /status world map. Used under CC-BY-4.0; only aggregated counts are stored, never individual IP addresses.
Bug reports, patches, and version corrections welcome via contact.
OWNERSHIP & SUPPORT
This service is operated by Greg Burd. I pay all hosting fees personally. No ads, no tracking, no data collection; the only cookies are the optional ones that remember your search settings and mail theme (privacy). This is a gift to the PostgreSQL community. Every page here is the same for everyone, and every archive is also a git repo you can clone and keep.
A Tip Jar is planned to help defray infrastructure costs. If you find this service valuable, please consider supporting it when it opens.
Questions? Get in touch.
Not affiliated with the PostgreSQL Global Development Group.