Exhibit

An AI data-analysis CLI that treats analysis as a persistent, resumable investigation — an append-only graph of typed steps (question → plan → read-only SQL → result → evidence-backed conclusion), not a chat transcript. Load a CSV/Parquet file (or snapshot a Postgres table), ask in plain English, and get an inspectable, replayable, exportable investigation.

Read-only by construction. Data is loaded once into a per-investigation DuckDB opened read-only with external access disabled; every generated query is parsed by a static guard that permits a single SELECT and blocks writes, ATTACH, COPY, and file access. See THREAT_MODEL.md.

Install

pip install exhibit-cli
export ANTHROPIC_API_KEY=...        # or run offline with --llm mock

Use

exhibit start data.csv                                   # one file
exhibit start orders.csv customers.csv                   # several (joins across them)
exhibit start postgresql://user@host/db --tables orders  # live Postgres (read-only snapshot)

Then ask questions in plain English. Inside an investigation:

• /judge — adversarial self-review (weak claims, missing evidence, untested alternatives)
• /branch <id> <question> — fork a question off an earlier step
• /export [html] — Markdown, or a self-contained HTML report
• /help — everything else

Resume anytime with exhibit open <id>.

License

Source-available under the Business Source License 1.1: read the source and use it freely in production, but don't offer a competing commercial or hosted data-analysis product. Converts to Apache-2.0 on 2030-07-31. Not an OSI open-source license.