Frequently asked questions

What does "grounded by construction" mean?

Before a word is written, draft mines your sources for claims. A claim survives only if its quote appears verbatim in the source and every number in it appears in that quote. The writer is given that verified ledger and nothing else, so a sentence with no backing claim cannot be written. See how grounding works.

How is this different from RAG, or from asking an LLM to summarise a paper?

Retrieval hands a model source text and hopes it stays faithful. draft verifies each claim before writing and drops anything unverifiable, then attributes every sentence to the claim that backs it. The check is structural, not a prompt.

Which agent CLIs does it use, and does it need an API key?

No API key by default. In auto mode draft drives whichever supported agent CLI you are already logged into — Claude, Copilot, Codex, Cursor, Grok, Gemini and more — through that tool's own session. Offline, it uses a local Ollama model.

If you have no agent CLI at all, --engine api:anthropic (or api:openai) is an opt-in escape hatch that calls a hosted API directly with your own key from the environment. It is never chosen automatically — the keyless path stays the default — and exists only so a machine without an agent CLI still has an online option.

Can I set defaults so I don't repeat flags?

Yes. Put them in a project draft.toml or a user ~/.config/draft/config.toml (engine, reader, models, output directory, style, signing key). Precedence is flags > environment > project file > user file > built-in default, so a config file never changes what a command already does. DRAFT_CONFIG names an explicit file; DRAFT_NO_CONFIG disables the layer.

Does my paper leave my machine?

It depends on the engine. With a local Ollama model, nothing leaves your machine. With a cloud agent CLI, the source excerpts draft needs to extract and write are sent through that tool, exactly as if you had pasted them into it — so choose the engine that matches your privacy needs. draft itself has no servers and sends no telemetry.

What runs fully offline?

--engine ollama keeps the entire pipeline local. Reading, claim verification, house-style checks and provenance are all deterministic Go and never touch the network.

What is the C2PA manifest, and what does --verify check?

Every article ships a per-sentence attribution file and a C2PA manifest. draft --verify recomputes the digests and reports whether the article still matches the ledger it was written from. Editing an unsupported sentence makes it fail. Configure a signing certificate (DRAFT_C2PA_CERT / DRAFT_C2PA_KEY, with c2patool installed) and draft also writes a signed .c2pa credential bound to the article, whose signature and trust chain --verify checks too. draft --verify --json prints a portable draft.verification-record/v1 receipt that any other tool can re-check.

Can draft check that a claim is used correctly, not just quoted?

The verbatim gate deliberately checks wording, not meaning. --second-gate closes that gap on request: a local model judges whether each verified quote actually supports its claim and drops the ones it does not. It is opt-in, strictly additive, and fail-open, so it can only tighten the ledger — see how grounding works.

What inputs are supported?

PDF, Markdown, plain text, DOCX and LaTeX (.tex) are all read directly. .docx is built in on macOS via textutil, and works on any platform with --reader docling; .tex needs no external tool and keeps a formula as exact text. PDFs are read by pdftotext by default, or by --reader docling when tables and structure matter — which also mines table cells into claims. The one remaining input limit is a scanned PDF with no text layer; OCR is on the roadmap.

Does it work with non-English papers?

The verification gate is largely language-agnostic: UTF-8 handling, normalisation and verbatim quote matching do not assume English. The English-biased parts are the house-style rules and sentence-boundary repair, which you can adjust with --style. Full locale-configurable style is on the roadmap.

Can I use the output commercially?

Yes. draft is free and open source under the MIT or Apache-2.0 licence. No account, no telemetry.