OPEN SOURCE · LOCAL-FIRST · AGPL-3.0

Trace every claim
back to solid ground.

Groundline is the trust layer between your documents, your data, and your AI agents - on your desktop, or across your team. Drop in the files behind a report and it finds the claims, links every number to its evidence, and flags what no longer holds.

Download for Windows · macOS (Apple Silicon) · Linux (.deb) - native app, not a browser in a trench coat. Deploying for a team?

Local-first.·Your files never leave your machine.·Works fully without an AI provider.

The questions every review quietly worries about -

EVIDENCE

Every number, traced to its cell.

Groundline detects the claims your documents make and matches each figure against the values in your spreadsheets and data files - exact and approximate matches honestly labeled, unsupported numbers called out.

  • Claims classified: observed results, forecasts, estimates, assumptions, causal claims
  • Numbers matched with row & column context, not just string equality
  • “Unsupported” is an answer too - silence never passes as evidence
QBR_Deck_Q2.pptx · SLIDE 3Supported · exact match
“Revenue grew 24% in Q2 2025.”
claim · observed result · detected in slide title
qbr_financials.xlsx · Revenue!D14formula · =D13/D9-1
Q2 growth, computed from source rows24.0%
QBR_Deck_Q2.pptx · SLIDE 9Unsupported
“Churn improved materially.”
no matching value in any workspace file
FINDINGS

Findings, not vibes.

Deterministic checks raise concrete, explainable findings - contradictions between documents, formulas silently replaced by typed values, charts drawing on hidden rows, causal language resting on correlation.

  • Every finding names its files, elements, and severity
  • Deterministic detectors - the same files produce the same findings
  • Macros detected and reported, never executed
WORKSPACE · QUARTERLY BUSINESS REVIEW4 findings
Contradiction between documentsSlide 3 reports 24% Q2 growth; the exec summary states 19% for the same metric.
Formula replaced by typed valueSummary!C9 contains a hardcoded 1,860,000 - neighbouring cells still compute =SUM(...).
Chart uses hidden rows“Pipeline by region” draws on 3 rows hidden in the source sheet.
Causal claim backed only by correlation“Marketing spend drove growth” - the workspace shows correlation (r = 0.83), not causation.
VERSIONS

New numbers. Old slides. Caught.

Upload a newer copy of a file and Groundline explains what changed in business terms - then flags every downstream slide, paragraph, and recommendation still carrying the old numbers, with a reviewable correction attached.

  • Plain-language version summaries, not cell-by-cell noise
  • Stale content flagged across every document in the workspace
  • Nothing changes without human approval
qbr_financials.xlsx · v2 → v3version diff
Q2 revenue growth  24.0%18.7%
Late-June credit memo restated two enterprise deals.
3 STATEMENTS NOW OUTDATEDneeds review
Deck · slide 3“Revenue grew 24% in Q2 2025.”
Report · ¶2“…continuing the 24% growth trend…”
Report · rec 1“Hire ahead of sustained 20%+ growth.”
PROPOSAL · slide 3 titleevidence attached
“Revenue grew 24% 18.7% in Q2 2025.”
✓ ApproveRejectView evidence
AI AGENTS

Agents on a leash.

An MCP server exposes the workspace to Claude, ChatGPT, or your coding agent as structured tools. Agents can trace numbers, validate claims, and propose corrections - every proposal lands in the human review queue.

  • Tools like groundline_validate_claim and groundline_trace_number
  • Provider citations validated against the workspace; fabricated references stripped
  • Optional external AI for the assistant - only claim texts and snippets are sent, never your files
agent session · MCP
◆ agent calls groundline_validate_claim
  { "claim": "Revenue grew 24% in Q2 2025" }

◇ groundline responds
  {
    "support":  "outdated",
    "evidence": "Revenue!D14 = 18.7% (v3, current)",
    "was":      "24.0% in v2",
    "proposal": "queued for human review"
  }

- no silent edits. ever.
1extract 2detect claims 3link evidence 4raise findings 5track versions 6propose fixes 7export audit report
FILE INTELLIGENCE

It reads office files the way an auditor would.

Not just the text - the formulas, the hidden rows, the speaker notes, the chart source ranges, the metadata. 29 formats, from an .xlsx to a .parquet to a scanned .png.

.xlsx / .xlsm / .xls / .ods

Formulas & cached values, hidden rows/columns/sheets, merged cells, charts and their source ranges, named ranges. Flags hardcoded formulas, broken references, totals that don't add up.

.csv / .tsv

Delimiter & encoding detection, per-column types and stats. Flags duplicate and near-duplicate rows, identifier collisions, category spelling drift, date gaps, suspicious rounding.

.pptx / .ppt / .odp

Slides, titles, body text, speaker notes, tables, chart data, hyperlinks - and embedded pictures via optional local OCR. Flags duplicate slides, truncated axes, charts without readable data.

.docx / .doc / .odt / .rtf

Headings, paragraphs, tables, hyperlinks, metadata - and embedded pictures via optional local OCR. Detects claims per paragraph and flags causal language for review.

.pdf

Page text and tables (best effort), metadata. Detects image-only pages and suggests local OCR - an on-device model reads them without the file leaving your machine; reports password-protected files as locked. Audit reports export back out as PDF.

.md / .txt

Headings, paragraphs, lists, tables, links. Claims and assumptions extracted like any other source.

.json / .xml / .parquet

Flattened into the same rows and columns the CSV path profiles - or field → value when the structure isn't tabular - so they get the same duplicate, collision and correlation checks.

.html / .eml

Converted to text with tables preserved as real tables. An email's subject line is analyzed too - it's very often where the headline number lives.

.png / .jpg / .webp / .tiff

Read with local OCR (Pro): nothing leaves your machine. Recovered text is labeled as OCR and its claims carry lower confidence, because it's a reading of a picture, not a value from a file.

Legacy Office and OpenDocument files are converted by LibreOffice and then read by the extractor for the modern format - one parser family, not four. If it isn't installed, the app says so by name instead of failing with a vague parse error.

LANGUAGES & NUMBERS

1.234,56 is not 1.234.

Claim detection reads English, Portuguese, Spanish, German and French, and numbers are read in both decimal conventions - so a European figure never arrives silently a thousand times too small.

  • Where a token is genuinely ambiguous (1.234), the document's own convention decides it - detected once per file
  • 27 currency codes and 12 symbols, written before the amount or after it (€1.234, 1.234,56 EUR)
  • Reporting periods in all five: Q3, T3, H1, FY 2026, 2. Quartal, 1º trimestre
  • The interface is still English - only the analysis is multilingual
Relatório_T3.docx · PTconvention: 1.234,56
“A receita subiu para 1.234.567,89 EUR no 3.º trimestre.”
claim · observed result · period Q3 · EUR 1234567.89
Q4_Bericht.docx · DEforecast
„Wir erwarten 12 % Wachstum im 4. Quartal.“
marker “wir erwarten” → forecast, not a result
WAS SILENTLY WRONG BEFOREfixed
1.234,56 → 1.2341.234,56 → 1234.56
three orders of magnitude, flowing into evidence matching as if it were real
CONNECTORS

The number lives in a warehouse. So go there.

Files don't only arrive by drag-and-drop. Watch a synced folder, read a query straight out of PostgreSQL, and re-check a workspace on a schedule instead of only when someone touches a file.

  • Watch folders are recursive - a synced SharePoint, OneDrive, Drive or Dropbox tree is nested by client and project, and a flat scan saw nothing
  • A data source is a named query; each refresh lands as a new version, so version diffs, evidence links and staleness apply unchanged
  • Every query runs in a READ ONLY transaction, capped by row count, with the connection string sealed at rest and never returned to the browser
  • Daily or weekly re-checks, so staleness surfaces on a Monday morning rather than at the board meeting
DATA SOURCE · revenue_by_regionpostgres · READ ONLY
refreshed 06:00 → qbr_financials.csv v4
Hand-exporting a CSV breaks the chain exactly where it matters: nothing records which query produced it, or when.
SCHEDULE · QUARTERLY BUSINESS REVIEWweekly
Mon 06:00re-ran the pipeline · 2 statements went stale
Watch folder/Clients/Acme/** · 3 new files imported
PRICING

For you. Or for the whole team.

The engine is open source and free forever. Pro is the polished desktop experience for individuals; Team adds reviewing together - shared workspaces, accounts, roles and single sign-on. Annual only - and license keys work offline: no accounts, no activation server, nothing phones home.

FREE · OPEN SOURCE

Groundline

$0 forever · AGPL-3.0

The full analysis engine on your machine: claims, evidence tracing, findings, version tracking, the offline assistant.

  • Every detector, all 29 formats, five languages
  • Watch folders, warehouse sources, schedules
  • Backup export and restore, encryption at rest
  • Build from source - it's all there
TEAM

Groundline Team

$350/seat/year · from 3 seats

Shared review on your own infrastructure - your git repository is the control plane, so data never leaves your network.

  • Everything in Pro
  • Git-backed shared workspaces
  • Accounts, roles, SSO and two-factor
  • One review queue, full audit trail
  • Onboarding with the founder
What's includedFREEPROTEAM
ANALYSIS - THE OPEN-SOURCE ENGINE
Claim detection & evidence tracing across 29 file formats
Five languages (EN · PT · ES · DE · FR) & both decimal conventions
Findings: contradictions, hardcoded formulas, hidden data, causal claims
Version tracking with plain-language diffs & stale-content flags
Correction proposals with human review
Offline assistant · local AI endpoints (Ollama, LM Studio…)
MCP server for AI agents
Watch folders, PostgreSQL data sources & scheduled re-checks
Workspace backup export & restore - your data is never hostage
Encryption at rest & optional local grammar suggestions
THE PACKAGED APP & DELIVERABLES
Signed installers & automatic updatesbuild it yourself
Audit report export (PDF or HTML)
Corrected copies of decks & documents (single + batch)
Local OCR for scanned pages, embedded pictures & image files
Hosted AI providers (OpenAI, Anthropic, ChatGPT subscription)
Priority support
COLLABORATION - WHEN IT'S SERVED TO A NETWORK
Git-backed shared workspaces & sync
Accounts, roles & per-workspace membership
Single sign-on (OIDC) & two-factor authentication
Team review queue, assignments by email, contributor graph & audit trail
Onboarding with the founder

Every Pro feature works free in the built-in demo workspace - a license is about your own files. Trials are the full product for 45 days, no card.
Accounts switch on only when the instance is served to a network host: a desktop or localhost install stays single-user, with no login screen.

OPEN SOURCE

Yours to run, read, and fork.

AGPL-3.0 - open source that stays open: nobody can sell a closed fork or a proprietary hosted copy. Local-first by design: a native desktop shell, a deterministic analysis pipeline, SQLite on your disk. A seeded demo workspace - with deliberately planted issues - shows every feature on first launch.

Local-first. Your files stay on your machine. The assistant works fully offline.
Deterministic. Same files in, same findings out. No model roulette in the pipeline.
Measured. A labelled corpus ships in the repo: 5 of 5 planted defects found, 0 false positives on the clean documents, across 9 cases - run it yourself with python -m benchmark.run. Nine cases at 100% means no regressions on what we thought to test, not perfection.
Auditable. Export a self-contained report: sources, claims, findings, versions, every AI proposal and its review decision.
~/code
$ git clone https://github.com/guilhermecariru/GroundLineOSS
$ cd groundline && npm install
$ npm run tauri dev
 analysis   pipeline ready · deterministic
 desktop    Groundline running · native window
 seeded     demo workspace “Quarterly Business Review”
   └─ planted: outdated deck, contradictory slide,
      hardcoded formula, duplicate records, causal claim…

Put your next review
on solid ground.

Free, open source, and honest about what it can't do. Install it, drop in the files behind your next report, and see what holds.

Direct downloads: Windows · macOS (Apple Silicon) · Linux (.deb) - new releases and changelogs: follow on GitHub.

Start your 45-day trial

Full Pro, no card. Your key is issued instantly, works offline, and carries your email as its license holder.

Trouble? Email us and we'll send a key by hand.