backdraft

Drop-in provenance for factual claims: one click from any claim to the receipt behind it.

An analyst who cannot show where a number came from either re-reads the source or does not use the output. Backdraft removes that cost. In the usual loop, an agent does the writing: it reads your source documents only through a gate that mints a citation token over every span it shows, cites those tokens inline, and a postprocess resolves them, attaches the verbatim evidence, and renders one self-contained HTML file, the memo with the actual cited pages and spreadsheet cells embedded inside it. The human just reads that file: click a claim, see the source.

It works the same whether the writer is a model or a person, and whether the citations were minted during writing or attached to a document afterwards.

Docs live at backdraft.dev, one page for people, and backdraft.dev/llms.txt for agents.

60 seconds

uv tool install backdraft
cd demo

Ingest the sources. Every anchor is minted here, chunks for PDF pages, cells for spreadsheet sheets. Formats: PDF, XLSX/XLSM, XLS, CSV/TSV, DOCX, PPTX, images (png, jpeg, tiff, through the vision model), plain text and Markdown.

$ backdraft init
registry: /…/demo/.backdraft
documents: 0

$ backdraft ingest sources/t12-summary.pdf sources/underwriting-model.xlsx
t12-summary  t12-summary.pdf  pdf  3 pages
underwriting-model  underwriting-model.xlsx  xlsx  2 pages
note: extracted with pdf-text (the embedded text layer). Glossy or scanned PDFs extract better through a vision model: set BACKDRAFT_VLM_API_KEY in .backdraft/env.

Read. The page arrives with a citable name over each chunk. This is the whole mechanism: what you can cite is exactly what you were shown.

$ backdraft read t12-summary p1
# t12-summary p1  (page 1 of 3)

[bd:t12-summary:p1.c1:c2e8]
Bridgeview Commons, Trailing Twelve Month Summary

Property: Bridgeview Commons, 4400 Halsted Avenue, Columbus, OH 43214. 128 units across four three-story
garden buildings on 6.2 acres, built in 1998 and substantially renovated between 2019 and 2021. This summary
covers the trailing twelve months ended March 31, 2026, prepared from the borrower's monthly operating
statements and reconciled against the property manager's general ledger. Figures are unaudited.

[bd:t12-summary:p1.c2:7f11]
Total effective gross income for the trailing twelve months was $2,684,400, against gross potential rent of
$2,972,160. The resulting economic occupancy of 90.3% reflects an average physical occupancy of 94.1% offset
by concessions, bad debt, and vacancy loss. Physical occupancy was 96.4% in the most recent month and has not
fallen below 92.0% in any month of the period.

… five more chunks, c3 through c7.

Chunks fall on the document's own paragraphs. A PDF stores glyphs at coordinates, not paragraphs, so the extractor rebuilds the breaks from the line geometry before the chunker sees the page.

Search results are citable too, no page read required to get an anchor:

$ backdraft search "24850000"
1 result for "24850000"

[bd:underwriting-model:assumptions!B10:964a]  underwriting-model p2
  24850000

[Read the page: backdraft read underwriting-model p2]

A cell you can already see mints directly:

$ backdraft cell underwriting-model "assumptions!B10"
[bd:underwriting-model:assumptions!B10:964a]  24850000

Cite by writing the token as the href of a markdown link on the claim span. Multiple citations are ;-separated in one href.

[net operating income of $1,429,600](bd:t12-summary:p1.c3:f10b)
[Effective gross income of $2,684,400](bd:t12-summary:p1.c2:7f11;bd:underwriting-model:rent-roll!B11:4b79)

Bind. Every token resolves or is reported. Nothing drops silently.

$ backdraft bind memo.md --session s-bridgeview --check value-trace,overlap
bound 15 claim(s), 16 citation(s) [frontwalk]
  resolved: 15
  unresolved: 1
  overlap: partial 4, pass 11
  value-trace: pass 15
  ! unresolved: bd:t12-summary:p4.c1:1a2b
wrote .backdraft/records/memo.backdraft.json

$ echo $?
2

Exit 2 means a citation did not resolve, the code a Stop hook or a CI job gates on.

Render. One self-contained HTML file: the document, the receipts, the evidence, cited page images, spreadsheet cells in place, the machine-readable record, all embedded, no network, nothing to install for the reader.

$ backdraft render memo.md --to html
memo.backdraft.html

That file is the whole deliverable: send it over email or Slack and the recipient gets the full experience by double-clicking it.

The full version of this, with every command and every output: demo/walkthrough.md. The artifact it produces is checked in at demo/memo.backdraft.html.

The design in five sentences

Reading is a gate: source documents reach the writer only through read and search, which mint a token over everything they show and record it in a ledger, so the set of citable things is exactly the set of things shown, and a citation to something the writer never saw is a distinguishable failure rather than an invisible one. Anchors are content-addressed: a token's identity derives from the document's content and a location inside it, and it binds to an extraction snapshot, so re-ingesting the same bytes yields the same tokens and a changed source yields drifted instead of a broken link. The receipt travels with the claim, an anchor is not a pointer but a pointer plus the verbatim snippet and its hash, which is why the finished artifact is defensible with the registry deleted and the sources gone. Verification is a set of switches, default off: resolution is inherent to binding, but value-trace, overlap and entail are opt-in, recorded as graded evidence and never used as gates, so out of the box this is provenance rather than a truth oracle. Artifacts are self-describing: each carries a $format string matched byte-for-byte and an embedded $legend that teaches a reader who has never heard of backdraft how to decode and check it, so the format outlives this implementation.

Failures are data throughout: unresolved, not_shown, drifted, malformed and unmatched are first-class records in the report and visible sections in the artifact. Nothing is ever warned about and dropped.

What lives where

A working directory stays clean: the authored document and its artifact are the only visible outputs. Everything else, registry, credentials, bind records, lives under one hidden .backdraft/ directory, and backdraft clean tidies strays from older versions. The registry contains the full text of everything ingested, so gitignore .backdraft/ for confidential corpora.

Install

uv tool install backdraft   # CLI, recommended form
uv add backdraft            # as a dependency
pip install backdraft       # or plain pip

Python 3.13+. The vision extractor ships by default, and it is the recommended path for real PDFs (glossy layouts, info boxes, scans): a vision model reads each page and its clean representation becomes the receipt. Installing it is not consent to use it. --extractor auto prefers it only on explicit, backdraft-scoped consent: BACKDRAFT_VLM_API_KEY (env or .backdraft/env, backdraft init writes a template). Ambient OPENAI_API_KEY/OPENROUTER_API_KEY are deliberately never read; a generic key in the environment is not consent to send documents to its provider. The default provider is OpenRouter running Gemini 3.1 Flash Lite; direct OpenAI is base_url + model, set explicitly. Without a scoped key, auto falls back to the keyless text layer and says so. PDF page rendering needs poppler, installed separately.

Two extras:

backdraft[vlm] still installs: it is an empty compat alias from when the vision deps were an extra.

From a checkout:

uv sync
uv run backdraft --help
uv run pytest

Skills

Three agent skills ship inside the package and in this repo:

The CLI is the system; a skill is one page telling an agent to use it. Nothing in the skills is required to use backdraft by hand. How the skills reach your agent depends on the harness:

Claude Code. The repo is its own plugin marketplace, so the plugin route tracks releases:

/plugin marketplace add spencerbraun/backdraft
/plugin install backdraft@backdraft

Or have the CLI copy the skills into your skills directory:

backdraft skill install          # the writing skill, into ~/.claude/skills/
backdraft skill install --all    # plus backfill and artifact-reading
backdraft skill install --project  # into this repo's .claude/skills/

Claude Cowork. Once the plugin is listed in Anthropic's community directory, it installs from Cowork's built-in skills directory. Until then, zip a skill folder from skills/ and upload it under Customize > Skills. Inside a Cowork session the CLI itself runs per-command as uvx backdraft ..., since installs do not persist between sessions.

Codex, Cursor, Copilot. Agents in this family read skills from ~/.agents/skills:

backdraft skill install --agent codex   # into ~/.agents/skills/

or commit the skill folders to your repo under .agents/skills/ so every checkout carries them.

Standing context, any harness. Paste this into your repo's AGENTS.md:

## backdraft (cited writing)
When a document must cite its sources, write it through the backdraft CLI:
it shows source text with a citation token over every span, and only shown
spans are citable. In a sandboxed session run every command as
`uvx backdraft ...` (no install, no PATH edits). Start from
`uvx backdraft --help`; ground truth is https://backdraft.dev/llms.txt.

Documentation

The three files under spec/ are the portable specification: another implementation reads them and nothing else.

Status

v1. The registry format, the token grammar and the artifact format are pinned by prose specs and golden-file tests. What is queued, and what is deliberately parked with the objection written down, lives in ROADMAP.md.

MIT licensed.