c3db

Query the C3DB competition-climbing warehouse from your terminal — as raw SQL or from your own AI coding tool. It's a thin client over C3DB's hosted, read-only MCP server: every query is a single SELECT/WITH, cost-capped, row-capped, and timed out server-side. Nothing about the vetted query core is reimplemented here.

pip install c3db
c3db login                                   # sign in via your browser (one time)
c3db query "SELECT * FROM event_standings WHERE final_rank = 1 LIMIT 5"
c3db connect claude                          # wire the warehouse into Claude Code

pip install "c3db[ask]"                      # natural-language querying (bring your own model)
export ANTHROPIC_API_KEY=sk-ant-…            # your own model key — c3db never sees it
c3db ask "who has won the most World Cup golds in lead?"

from c3db import Client

df = Client().query("SELECT athlete_name, count(*) wins "
                    "FROM event_standings WHERE final_rank = 1 GROUP BY 1 ORDER BY 2 DESC")

Commands

Natural language — bring your own model

c3db ask and Client().ask() are bring-your-own-model: your own LLM reads the warehouse schema and writes the SQL; c3db validates and runs it. c3db never sees or pays for your model key — it's separate from the c3db_… data key. Install a provider and set its key:

pip install "c3db[anthropic]"    # or "c3db[openai]", or "c3db[ask]" for both
export ANTHROPIC_API_KEY=…       # Anthropic (default: claude-sonnet-5 — plenty for NL→SQL)
export OPENAI_API_KEY=…          # OpenAI-compatible; add C3DB base url / --base-url for DeepSeek/GLM/local
c3db ask "top 3 nations by lead World Cup wins since 2020"

Already use Claude Code? Then Claude needs no key at all — if you've run ant auth login (or Claude Code is set up), c3db ask picks up that OAuth login automatically. Just pip install "c3db[anthropic]" and ask.

Provider auto-detects (Anthropic by default — including via that OAuth login; override with --provider); --model picks the model. Client().ask(q) returns a DataFrame with the generated SQL on df.attrs["sql"].

The two kinds of access

• Raw SQL (c3db query, Client().query()) — you write the SQL; c3db runs it and hands back rows.
• Your own AI (c3db connect) — your AI tool's model reads the schema and writes the SQL itself, then runs it through the same server. This is bring-your-own-model: c3db never sees or pays for your AI subscription; it only issues the c3db_… data key that identifies you to the warehouse.

Results, limits & bulk data

A query returns up to 50,000 rows to the client (add a LIMIT for more than the default 1,000; the scan is capped at 2 GB and 60 s regardless). Save results locally as csv / parquet / json. The AI path (c3db connect) only sees a ~50-row preview plus the true row count — your model answers from that and points you here for the full data. For the whole raw dataset (bulk), self-serve download is coming; for now email yvonnelau.work@gmail.com.

Config

Resolved in order: environment → ~/.config/c3db/config.json → built-in default.

c3db login mints your key self-serve (sign in on the C3DB site → the key is created and stored locally). No key is typed at the terminal. Manage or revoke keys any time at the site's login page.