receipts

Source-verified research briefs

14.4 s · $0.15 · 19 supported claims (live)

crates.io · receipts 0.2.1

Overview

Ask a question, get back claims you can cite — each one pinned to a URL and a verbatim quote from the source. Anything that can’t be verified lands in a structured uncertainties field instead of being asserted as fact. Built for agents that need ground truth before writing a claim into a document.

ask a model directly

A confident paragraph of prose. No link, no quote — take it or check it yourself.

ask receipts

claim: the statement, in plain language
verdict: supported
source: https://example.org/…
quote: “…the exact sentence it came from…”

how a question becomes a cited brief

question inYou ask, in one line — agent or human.
fan outDecomposed and searched — parallel Exa queries, not one.
read & gateSources fetched and read. A relevance gate drops off-topic claims before verification.
verifyEach claim is checked against a verbatim quote in the source — not trusted on the model’s say-so.
cited briefEvery claim carries a URL and a quote. What can’t be pinned lands in uncertainties instead.

14.4 s · $0.15 · 19 supported claims (live)

Install

Homebrew

brew install treygoff24/tap/receipts

Shell installer — prebuilt binary, no toolchain needed

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/treygoff24/receipts/releases/latest/download/receipts-installer.sh | sh

Cargo

cargo install receipts

Configure

Set the key, run the doctor, done.

CEREBRAS_API_KEY required: inference — get one at cloud.cerebras.ai
EXA_API_KEY required: web search — get one at exa.ai

verify your setup

receipts doctor --online

How it works

Decomposes your question and fans out parallel Exa web searches, then reads the returned sources.
A verification pass re-reads each source before a claim earns a supported verdict — quotes are checked, not trusted.
Every claim carries both a verdict (claim vs. source) and a relevance (claim vs. question); a relevance gate runs before verification so off-topic claims never get marked supported.
Depth tiers scale worker count: quick ≈ $0.05–0.10, standard ≈ $0.15, deep ≈ $0.31 per question.
Budget caps (--max-dollars, --max-seconds) drain in-flight work and return a usable partial envelope instead of crashing.

Pairs well with

scoutthe same verification discipline pointed at local files instead of the web exa-agentthe search API receipts drives internally, exposed at full depth

Known limits

Can’t see behind logins or paywalls (e.g. PACER dockets) — those come back partial or unanswered, honestly labeled.
No synthesized prose answer by default: data.claims is the API surface; --brief adds an explicitly non-citable narrative.