draft
A printed research paper on a desk

Features

Features

Grounded, keyless, provable. What draft does, and how it proves it.

Grounded by construction

A small local model will invent a plausible number. A cloud API will charge for it and want a network. draft takes neither risk: before a word is written, your sources are mined for claims, and a claim survives only if its quote appears verbatim in the source and every number in it appears in that quote.

What it does

  • Any agent you already have. Ten CLIs supported — Claude, Codex, Copilot, Cursor, Grok and more — driven headlessly through their own logged-in sessions. No API key. For a machine with no agent CLI, --engine api:anthropic (or api:openai) is an opt-in escape hatch that reads your own key from the environment — the keyless agent-session path stays the default and is never chosen for you.
  • Works offline. When a session call fails because you are on a plane, the chain advances to a local Ollama model and stays there.
  • Verbatim grounding. Quote-checked claims, numeric cross-checks and metric-conversion detection. Unverifiable claims are dropped before writing.
  • A semantic second gate, on request. --second-gate adds an opt-in pass in which a local model judges whether each verified quote actually supports its claim, dropping the ones it does not. It is strictly additive — off by default, fail-open, and the verbatim gate stays primary — closing the one gap the verbatim gate is candid about: it checks wording, not meaning.
  • Claims from tables, not just prose. With --reader docling a table is preserved as Markdown, and draft mines each numeric cell into a grounded claim — a value with its row and column headers — verified against the source like any other.
  • Every input format. PDF, Markdown, plain text, DOCX, and LaTeX (.tex) — read directly, no external tool, with a formula kept as exact text where pdftotext would scramble it. pdftotext reads a 62-page paper in ~110 ms; --reader docling adds tables and structure when they matter.
  • House style, enforced. Banned words and phrases, British English, sentence-rhythm and structure rules — checked, not merely requested. Configurable with --style.
  • Provenance you can check — and sign. Per-sentence attribution plus a C2PA manifest beside every set; draft --verify recomputes the digests. Configure a signing certificate (DRAFT_C2PA_CERT / DRAFT_C2PA_KEY, with c2patool installed) and draft emits a signed .c2pa credential bound to the article — whose signature and trust chain --verify then checks as well.
  • A portable verification record. draft --verify --json emits a self-contained draft.verification-record/v1 receipt — the article digest and whether it matches, the grounding summary, the signature state, and the verdict — that any other tool can re-check. Bring your own artifact, verify anywhere.
  • Configure once. A project draft.toml or a user ~/.config/draft/config.toml sets your defaults — engine, reader, models, output directory, style, signing key. Precedence is flags > environment > project file > user file > built-in default, so nothing about an existing setup changes.
  • Publish-ready sets. Body, frontmatter and combined document, written side by side and regenerable without losing a curated field.
  • Never re-pay for extraction. A failed run leaves its verified ledger on disk; --resume re-verifies it and skips straight to writing.

Grounded, keyless, provable

Point it at one paper or twenty. Each becomes its own draft, and every draft can prove itself against its sources.