A desktop application for real statistical work — 239 procedures across Basic Statistics, Regression, ANOVA, Calc, Data and Graph — that runs entirely on your own machine. No subscription. No account needed. Your data never leaves your machine.
Named for William Sealy Gosset, who published the t-distribution as "Student" in 1908 — because his employer would not let him publish under his own name.
Gosset is a Minitab-equivalent analytics application. You open a worksheet, choose a procedure from the menu bar, fill in a dialog, and get a result window with tables, findings and graphs — then export the session as a typeset PDF, a Word document, an Excel workbook, a PowerPoint deck or Markdown.
Three ideas run through it:
Every result is a document, not a data dump. A procedure returns a narrative conclusion, headline statistics, titled tables and graphs. The PDF export is a designed report — a cover block, one bordered card per result, a coloured verdict badge, small-caps group labels, monospaced numerals, and the t-distribution rule on every page.
Power without clutter. A dialog with fifteen settings opens looking like it has four; the rest live in a collapsed Options section. Buttons say what they do — "Run regression", never "Submit".
Honest about its methods. Where an algorithm is proprietary, Gosset substitutes an equivalent and says so in the interface: Shapiro-Wilk in place of Ryan-Joiner, Anderson-Darling p-values from D'Agostino & Stephens, Bonett's p by bisecting its own confidence interval, Dixon's Q against published tables for n ≤ 30.
 |
 |
| A result window — light… | …and dark. Both themes are first-class. |
| Menu | What's in it |
|---|---|
| File | Open (CSV, Excel, JSON, Google Sheets); save and open .gsp projects carrying every worksheet, fitted model, constant, matrix and formatting rule; Export Report to PDF / Word / Excel / PowerPoint / Markdown; Print; Options |
| Edit | Per-worksheet undo/redo, clipboard, cell and range operations |
| Data | 25 operations in Minitab's own order — Subset, Split, Merge, Stack Worksheets, Sort, Rank, Delete Rows, Erase Variables, Copy, Stack/Unstack/Transpose, Recode, Change Data Type, Date-Time, Concatenate, Conditional Formatting, Display Data, Worksheet Information, Constants |
| Calc | 28 procedures, 78 dialogs — the Calculator (an ast-whitelisted expression engine over 49 functions, with live validation and a caret under the offending character), Column/Row Statistics, Standardize, Make Patterned Data, Make Mesh Data, Make Indicator Variables, Set Base, Random Data and Probability Distributions (26 distributions in Minitab's parameterisation), Resampling (5 bootstrap and randomization tests), Matrices |
| Stat › Basic Statistics | All 18 — 1-Sample Z and t, 2-Sample t, Paired t, 1 and 2 Proportions, 1 and 2-Sample Poisson, 1 and 2 Variances, Correlation, Covariance, Normality Test, Outlier Test, Poisson Goodness-of-Fit, Descriptive Statistics, Store Descriptives, Graphical Summary |
| Stat › Regression | All 13 — Fitted Line Plot, Fit Regression Model, Predict, Best Subsets, Stepwise, Nonlinear, Orthogonal, Partial Least Squares, Stability Study, Binary/Ordinal/Nominal Logistic, Poisson Regression |
| Stat › ANOVA | All 20 — One-Way (both layouts, Welch, Tukey/Fisher/Dunnett/Games-Howell with grouping letters), Test for Equal Variances, Balanced ANOVA, Fully Nested ANOVA, General MANOVA, a General Linear Model submenu (fit → comparisons, predict, factorial plots, contour, surface, response optimizer), a Mixed Effects Model submenu, interval/main-effects/interaction plots, Analysis of Means |
| Stat › more | Tables (chi-square), Time Series (forecasting), Multivariate (segmentation), plus decision tree, random forest, gradient boosting and AutoML |
| Graph | Scatter, histogram, boxplot, interval, matrix, contour, surface and 3-D plots — interactive by default, switchable to static in one place |
| Window | The MDI window manager, the Session Window, the Report pane, the Icon Gallery |
Also: a two-row Minitab-style worksheet header, multiple worksheets as tabs with per-sheet undo,
constants (K1…) and matrices (M1…) with their own windows, conditional formatting that recomputes
on every edit, and a Report pane for curating a report rather than dumping a whole session into one.
Every menu item carries an icon and a hover help card saying what it does and what data it needs; Window › Icon Gallery shows the whole set.
Run the installer and you're done. Once installed, Gosset keeps itself up to date — it checks for new releases in the background and offers to install them, so this is the only time you need this page.
Gosset installs per-user — no administrator rights, no UAC prompt. It adds a desktop shortcut and a
Start Menu entry, and registers .gsp files so double-clicking a project opens it.
The installer is not code-signed. A code-signing certificate is a paid annual subscription and Gosset does not have one. So the first time you run the installer, Windows SmartScreen will show:
Windows protected your PC — Microsoft Defender SmartScreen prevented an unrecognized app from starting.
To continue, click More info, then Run anyway.
That message means "Windows has not seen this file signed by a publisher it knows" — not "this file is known to be harmful". It will appear for every release until there is a certificate. If that trade-off isn't acceptable to you, build from source instead (below); the instructions produce the same application.
Everything runs locally — an ordinary Windows application with a bundled Python analysis engine on
127.0.0.1. No telemetry, and no account required.
The only two things that ever touch the network, both optional and both explicit:
- Update checks, which ask GitHub whether a newer release exists. Turn them off in File > Options > Updates.
- Signing in with Google, if you choose to and if the build has it configured. It puts your name in the menu bar and on report covers — nothing else. Your data is never uploaded, no feature is gated behind it, and the app is fully usable signed out. See docs/FIREBASE_SETUP.md.
Your worksheets, projects and reports stay on your machine either way.
Requirements: Windows 10 or 11, 64-bit; about 700 MB on disk (the bundled scientific Python stack is most of that).
Gosset checks GitHub for a new release a few seconds after starting, and every four hours after that. When one exists it shows the release notes and asks — it never downloads anything on its own.
- Update now downloads the new version with a progress indicator, then offers to restart. If the session has unsaved work, it asks about that before restarting.
- Later stops asking until the next time you start Gosset.
- A failed check is silent. Offline, behind a proxy, or GitHub unreachable: Gosset says nothing and carries on, because a dialog every launch about a thing you cannot fix is worse than no dialog.
Turn it off in File > Options > Updates, which also has a Check now button and shows the version you are running. After an update, a What's new window shows the notes for every version you skipped, once.
Uninstalling: Apps & features → Gosset → Uninstall, which removes the app, its shortcuts and the
.gsp association. Your saved projects and exported reports are left alone, and so is
%APPDATA%\Gosset (window position and preferences) — delete that folder by hand for a completely
clean slate.
The PDF comes from a report engine that knows no statistics at all — it is handed titles, rows, a PNG and a caption. One translator module knows the shape of a result; the engine only knows typography. That split is what makes a new procedure need no engine change.
Its corollary: a verdict badge is computed from numbers — p_value, r_squared, accuracy — and
never parsed from prose, so rewording a conclusion cannot flip a badge from red to green. Polarity
means attention, not good news: a significant normality test is red, because the assumption
failed.
Charts are re-rendered for export at 2× in the light palette, so a report exported from dark mode still has light figures, typeset for paper rather than for a 300-pixel panel.
┌───────────────────────────────┐
│ backend/core/ + the │ pandas · scipy · statsmodels
│ report engine │ scikit-learn · matplotlib · reportlab
└───────────────┬───────────────┘
│ ONE implementation
┌───────────────┴───────────────┐
│ │
┌────────▼─────────┐ ┌──────────▼─────────┐
│ FastAPI REST │ │ MCP server │
│ + browser UI │ │ (10 tools, stdio) │
└────────┬─────────┘ └──────────▲─────────┘
│ │
┌────────▼─────────┐ Claude and other MCP
│ Electron shell │ clients drive the same
│ (the .exe) │ analyses conversationally
└──────────────────┘
The analysis code has two front doors onto one implementation, and that is the genuinely unusual part:
- A REST API with a browser UI. The frontend is plain ES modules and CSS with no build step — no
bundler, no transpiler, no
node_modules— served by FastAPI itself. The desktop app is this same UI in an Electron window. - An MCP server.
python mcp_server.pyexposes 10 tools over the Model Context Protocol, so Claude (or any MCP client) can load a dataset, run a regression, forecast a series and export a report by being asked to. Those tools call the same functions the dialogs call — not a reimplementation, and not a robot driving the UI.
Inside the desktop app, Electron picks a free localhost port, starts the bundled Python backend as a
sidecar, waits for /health, and only then opens a window — so the first thing on screen is the
worksheet, not a connection error. The backend is terminated on quit, and watches its parent from its
own side so a hard-killed shell cannot orphan it.
A few rules that shaped the codebase, if you're reading it:
- One endpoint per menu area, not per procedure.
POST /datasets/{id}/anovatakes aprocedurestring and dispatches through a handler dict; 20 request models would be pure boilerplate. - Adding a procedure is one config entry plus one backend handler. A shared dialog builder owns every form and the standard result layout.
- The Calculator never evaluates a string. It parses with
ast.parseand walks the tree with one visit method per allowed node type — a whitelist by construction, not a blacklist of things to strip. - A fitted model is client state. The GLM and Mixed dialogs return a model spec that downstream dialogs post back for a refit, so the API stays stateless and a fitted model travels inside a saved project.
DESIGN.md documents the visual language — tokens, type, motion, the report surface and the
ban list — and CLAUDE.md documents the architecture and the traps in full.
Requirements: Python 3.11, Node.js 20+, and Windows if you want the installer.
git clone <this repo>
cd personal-analytics-mcp
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txtuvicorn backend.api:app --reload --port 8000Open http://localhost:8000. This is the whole application; the desktop shell wraps it and never
replaces it. LaunchBackend.bat does the same and opens the browser for you.
python mcp_server.py # stdio — point an MCP client at itcd desktop
npm install
npm run build:sidecar # PyInstaller freezes the backend (~10 min, ~265 MB)
npm startnpm start works before build:sidecar too — with no frozen bundle present the shell falls back to
running the backend out of .venv, which is far faster to iterate on.
cd desktop
pip install -r ../requirements-build.txt
npm run build # sidecar, then the NSIS installerIt lands in %LOCALAPPDATA%\gosset-build\electron. Build artifacts go outside the repository on
purpose — see the comment in desktop/scripts/paths.mjs.
One-time prerequisite on Windows. electron-builder downloads a code-signing toolchain even though
this build is unsigned, and that archive contains macOS symlinks it cannot extract without symlink
privileges — the build then fails with Cannot create symbolic link ... libcrypto.dylib. Seed the
cache once, without the part a Windows build never uses:
$cache = "$env:LOCALAPPDATA\electron-builder\Cache\winCodeSign"
New-Item -ItemType Directory -Force $cache | Out-Null
Invoke-WebRequest -UseBasicParsing -OutFile "$cache\winCodeSign-2.6.0.7z" `
'https://github.com/electron-userland/electron-builder-binaries/releases/download/winCodeSign-2.6.0/winCodeSign-2.6.0.7z'
& 7z x "$cache\winCodeSign-2.6.0.7z" "-o$cache\winCodeSign-2.6.0" '-xr!darwin'(The release workflow does this itself. Enabling Windows Developer Mode also works, since that grants the privilege.)
python desktop/scripts/smoke_sidecar.py http://127.0.0.1:8000Thirteen checks over HTTP, each picked to be the first thing that drags a whole library in: pandas,
scipy, statsmodels, patsy, scikit-learn, matplotlib, reportlab, svglib, python-docx, openpyxl,
python-pptx. It exists because PyInstaller happily reports success for a bundle whose statsmodels is
missing a hidden import — a failure that only appears when the code path actually runs. The release
workflow runs it against the frozen backend before building the installer.
There is no automated test suite. Verification is three layers, in order: a scipy/statsmodels cross-check in Python, then the live REST API, then the real dialogs driven in a browser. The browser layer has repeatedly caught bugs the other two cannot.
desktop/package.json's version is the single source of truth. npm run stamp propagates it to
backend/version.py (report footers), frontend/brand/version.js (the About window) and
frontend/changelogData.js (the in-app "What's new" notes, generated from CHANGELOG.md). All three
are committed, so a source checkout needs no Node.js to run, and npm run stamp -- --check fails if
they have drifted.
One command ships a release:
cd desktop
npm run release -- patch # or minor / major / an explicit 1.4.2It bumps the version, inserts a CHANGELOG.md section for you to fill in, re-stamps, commits, tags
v<version> and pushes. Pushing the tag is what starts the build. Add --dry-run to see every step
without changing anything.
The workflow then builds the installer, publishes it to a GitHub Release together with latest.yml and
the .blockmap — the two files the in-app updater needs — and replaces the release body with that
version's CHANGELOG.md section. One changelog, three surfaces: the file, the Releases page, and the
app's "What's new" window.
The release is refused before anything is built if the tag disagrees with package.json, if the
changelog section is missing or still holds its placeholder, or if the update feed was not produced.
Source-available, all rights reserved — see LICENSE.
You may read the source, build it, and run it for personal, internal or educational use. You may not redistribute it, sell it, or offer it as a service. The name Gosset, the wordmark and the G-mark are not licensed.
Bundled third-party components keep their own licenses, including five typefaces under the SIL Open
Font License (backend/report_engine/fonts/).
Gosset performs statistical computation. Its results are not certified for regulatory, clinical, safety-critical or financial use — verify anything you rely on.