Everything Vestiga does, explained.
A short, navigable map of every module — hotkeys, workflows, and the data that lives on your machine. Nothing here leaves your PC.
Vestiga is a desktop context layer for Windows. One palette gathers your clipboard history, screenshots with OCR, structured extraction, file renaming, value detection, and a colour picker — all running locally on your machine via Ollama.
Open the palette with Alt+Space, cycle modes with Tab, or jump straight to a module with its direct hotkey. Every hotkey is rebindable in Settings → Hotkeys.
OCR runs on glm-ocr. Text tasks (summary, tasks, key-points, tables, filename suggestions) run on qwen3:4b. Both models are pulled to your machine at first run and never leave it.
Settings → Privacy & data.Command palette#
Alt+SpaceThe single window through which everything is reached. Opens, morphs between modes, closes when it loses focus or on Esc.
- Type to fuzzy-filter the command rows.
- ↑ / ↓Navigate rows.
- EnterRun the highlighted command.
- TabCycle modes — Commands → Clipboard → Screenshots → Convert → Commands.
- EscStep back one level, then close.
- Ctrl+PPin the highlighted clipboard entry.
- Ctrl+D / DelDelete the highlighted entry.
Clipboard history#
Ctrl+Shift+VEverything you copy — with the source attached. Foreground app title, plus browser URL when copied from a browser, or document path when copied from a PDF, DOCX, Excel, Notepad, and so on.
Row actions (icons on hover)
- Copy back — writes the entry to the system clipboard.
- Open source — launches the URL in the browser, or opens the source file in its native app.
- Show in folder — opens Explorer with the source file selected.
- Ctrl+PPin — pinned items sort to the top and survive history trim.
- Feature (bookmark icon) — separate flag with its own “Featured” tab.
- DelDelete.
Filter bar: All · Text · Link · Email · Phone · Address · Number · Featured. Each content type can also be toggled in Settings → Clipboard so it's never written to history at all.
Consecutive duplicate copies don't create duplicate rows (configurable).
Screenshots + OCR#
PrtScCapture a region, copy it to the clipboard, and let OCR run in the background. Local, fast, non-stealing.
- Press PrtSc → the full virtual screen is frozen in memory.
- An overlay appears showing the frozen frame; drag a rectangle over it.
- The cropped image is copied to the clipboard and saved to
%LocalAppData%\Vestiga\Screenshots\(format configurable: PNG / JPG / WebP). - A small toast appears in the bottom-right; OCR runs in the background and the toast's spinner turns into a green check when text is ready.
- Click the toast → full preview window opens with the OCR result.
Preview-window tools
- OCR text — extracted text, editable. Corrections are saved back to the DB; the feedback loop learns frequent substitutions and auto-applies them to future captures.
- Summary — one-sentence topic + 3–5 bullet facts, generated locally on demand.
- Copy — copies the active tab's text.
- Open source / Show in folder — same source-tracking as clipboard.
Screenshot history#
Ctrl+Shift+HThumbnail list of every saved capture with its OCR text and source. Full-text search runs against OCR content (SQLite FTS5). Same row actions as Clipboard, plus Delete.
Extract structure#
Ctrl+Shift+EStructured extraction over selected text or an entire document — summary, tasks, key points, tables.
- With text selected in another app — the hotkey reads the selection via UI Automation, opens the Extract window, and returns summary, tasks, key points, and any detected tables.
- With no selection — the palette opens in Extract mode so you can drag a file (PDF, DOCX, TXT, RTF, Markdown, code) onto it. PDF and DOCX are parsed locally, then the LLM extracts structure.
qwen3:4b. Context window is 8K — works for ~30–35K characters of mixed Cyrillic/Latin. Longer inputs are truncated from the middle with a warning row.Smart rename#
Ctrl+Shift+ROr right-click → "Rename with Vestiga". Select one file in Explorer, press the hotkey — Vestiga reads a snippet of the content, asks the local LLM for three descriptive names, and presents a chooser.
- Click a suggestion → file renamed.
- Type your own → rename to the typed name.
- Extension is preserved automatically (you edit the stem only).
The context-menu entry is registered during install. It also works from Windows 11's "Show more options" submenu for multi-select.
Convert#
Ctrl+Shift+COr inside the palette: Ctrl+C. A manual conversion surface. Type an amount, ticker, address, or time, and matching detectors fire.
- Currencies —
$200,1500 EUR,2.5k RUB→ converted to your base currency via open.er-api.com. Base currency is auto-picked from your Windows region on first run, changeable in Settings. - Time zones —
6pm EST,14:00 MSK→ shown in all common zones. - Stock tickers —
AAPL,TSLA→ live price (toggleable). - Crypto tickers —
BTC,SOL→ CoinGecko price. - Crypto addresses — BTC / ETH / SOL / TRON → token metadata via DexScreener.
- Transaction hashes — opens the right block explorer.
- Colours —
#FF6347,rgb(255, 99, 71),hsl(9, 100%, 64%), or CSS names (tomato,deeppink, …) → resolved to hex.
Any detector can be turned off in Settings → Detectors, and the external APIs have their own off-switches in Settings → Privacy & data.
Detect & convert#
Ctrl+CAny time you press Ctrl+C on a single-line selection that is entirely made of a detectable value, a small overlay appears near the cursor with the normalized result — no palette needed.
- Copy
$200→ overlay shows 200 USD → 16,120 ₽. - Copy
6:00 AM EST→ overlay shows the time in your local zone. - Copy
0x1234…beef→ overlay shows the token + its price. - Copy
#FF6347→ overlay shows a filled swatch of that colour + hex.
Enter copies the highlighted result, Esc dismisses.
Pick colour#
Ctrl+Shift+PEyedropper. Press the hotkey, cursor becomes a crosshair, click any pixel anywhere on the desktop.
- Filled 32×32 colour swatch.
- Hex value in Consolas (
#E4E4E4). - Copy button.
Esc or right-click cancel. The sampled pixel comes from a frozen snapshot taken at the moment you invoke the picker — so animated UI stays stable.
Typing in the palette with no mode selected searches across clipboard, screenshots, and extracts at once. Results show which module they came from and jump to the right mode on Enter.
Opened via the gear icon in the palette header, or by typing settings.
- General — start with Windows, minimize to tray, close behaviour.
- Hotkeys — every hotkey is rebindable; conflict detection blocks save until resolved.
- Clipboard — content types to capture, dedup, history size.
- Screenshots — format, JPEG quality, OCR language, output folder, auto-copy-to-clipboard toggle.
- OCR learning — enable/disable the corrections feedback loop, export/import learned substitutions.
- Detectors — per-detector on/off, base currency picker.
- Privacy & data — per-API network toggles, full data export (zip), clear clipboard / screenshot history, reset settings.
- About — version, links, update status, licences.
Everything Vestiga writes lives under one folder.
%LocalAppData%\Vestiga\settings.json— preferences.%LocalAppData%\Vestiga\workflow.db— SQLite (clipboard, screenshots, OCR corrections, actions).%LocalAppData%\Vestiga\Screenshots\— image files.%LocalAppData%\Vestiga\cache\— HTTP response cache for external-API detectors.
Vestiga runs on your machine. We don't have servers, we don't see what you copy, screenshot, or extract, and we can't see what you type. Everything on this page is verifiable — the source is open, and the full list of outbound network calls is enumerated below.
What stays on your machine#
- Clipboard history —
%LocalAppData%\Vestiga\workflow.db - Screenshots —
%LocalAppData%\Vestiga\Screenshots\ - OCR text — inside the same SQLite database.
- Extracts (structure / summary output) — same database.
- OCR corrections — learned substitutions, same database.
- Settings —
%LocalAppData%\Vestiga\settings.json - HTTP response cache —
%LocalAppData%\Vestiga\cache\(detector responses only).
None of this is ever transmitted to Vestiga or to a third party by us. The only time a byte of this data leaves your computer is if you click Copy, Share, or paste into another app.
AI processing#
OCR and LLM inference run locally via Ollama on your CPU/GPU. The models (glm-ocr, qwen3:4b) are downloaded once during first-run setup from Ollama's public model registry (registry.ollama.ai). After that first download, no further model traffic is needed.
When you…
- Take a screenshot → the image is sent to the local Ollama HTTP server at
http://localhost:11434, OCR'd in-process, and the result written to the local SQLite database. The image itself never leaves your machine. - Summarise / extract structure → the OCR text (or the file contents) are sent to the local Ollama HTTP server.
localhostonly. - Smart-rename a file → a snippet of the file's content is sent to local Ollama. Same story.
Outbound network calls (the complete list)#
These are the only hostnames Vestiga ever contacts. Every one of them is toggleable in Settings → Privacy & data — turn them off and the feature that depends on them falls back to its offline behaviour (usually “skip”).
All calls are plain HTTPS GETs with no authentication, no user identifier, no cookies, no tracking pixels. The URL path includes just the symbol / address / amount being queried (e.g. api.coingecko.com/…/simple/price?ids=bitcoin). Responses are cached locally so the same query doesn't re-hit the network for the duration of the cache TTL.
What we don't do#
- We don't have analytics, telemetry, or crash reporting that phones home. Debug logs stay on your machine and are never uploaded.
- We don't read your clipboard outside of the user-invoked paths (the clipboard listener ignores everything while the auto-detect gate is disabled).
- Paid tiers don't change any of the above. Upgrading unlocks additional features, it does not grant us access to your data — the product keeps running locally regardless of tier.
- We don't share with advertisers — we have no advertisers, because Vestiga is a local product and doesn't need any.
Third-party services#
External APIs (CoinGecko, DexScreener, open.er-api.com, Ollama's registry) have their own privacy policies. When you hit them you're hitting them directly from your IP — Vestiga doesn't proxy. Each service sees:
- Your IP address.
- The URL path (symbol / amount / model name).
- A generic
User-Agentheader.
No Vestiga-specific identifier is added.
Deleting everything#
Close Vestiga, then:
rmdir /s "%LocalAppData%\Vestiga"This wipes clipboard history, screenshots, OCR text, extracts, corrections, cache, and settings. Uninstalling Vestiga via Add or Remove Programs removes the binary but leaves this folder alone (so re-installs don't lose history) — you have to delete it yourself if you want a true clean slate.
The LLM models live under Ollama's data directory, not under Vestiga's.
ollama rm qwen3:4b-instruct-2507-q4_K_M
ollama rm glm-ocrSymptoms you might hit on a fresh install, grouped by module. If none of the fixes here work, file an issue at the link on the About page with a short description and the contents of %LocalAppData%\Vestiga\settings.json.
Install / first run#
Expected. The AI-engine install step (Ollama) is silent and takes 1–3 minutes on most machines while the engine unpacks. You don't have to do anything — the bar will finish on its own. Wait up to 5 minutes before assuming it hung.
The wizard pulls the two LLMs (~2.2 GB vision + ~2.5 GB language) from Ollama's CDN on first launch. On a slow connection this can take 20+ minutes. Bandwidth is the only limit; no Vestiga server is involved.
If it truly hangs, open a terminal and run:
ollama pull qwen3:4b-instruct-2507-q4_K_M
ollama pull glm-ocrRestart Vestiga — the wizard detects the models and advances.
We're shipping without EV code signing for 0.1. SmartScreen flags any unknown publisher. Click More info → Run anyway. This will go away once code signing lands in 0.2.
Hotkeys#
Something else on your machine already owns Alt+Space (often the language switcher on Windows 11). Vestiga shows a yellow warning in Settings → Hotkeys if a registration failed. Change the palette hotkey to something free like Alt+Shift+Space.
Windows 11 routes PrtSc to Snipping Tool by default. The Vestiga installer flips HKCU\Control Panel\Keyboard\PrintScreenKeyForSnippingEnabled = 0 if you left the corresponding checkbox ticked during install.
Manual fix if you unticked it:
Settings → Accessibility → Keyboard- Turn off "Use the Print Screen button to open screen snipping"
- Sign out and back in.
The context-menu entry is registered post-install via a CLI call. Rare install races can miss it. Restore it manually from an elevated prompt:
"C:\Program Files\Vestiga\WorkflowAI.exe" --install-shell-extOn Windows 11 the entry may live under "Show more options" rather than the top-level menu — that's standard Windows behaviour, not a bug.
OCR / screenshots#
Local inference on the default qwen3:4b + glm-ocr pair needs roughly 6+ cores and 12+ GB RAM to feel snappy. On 4-core / 8 GB machines the first inference of a session stalls while models load, and subsequent ones are steady-state CPU-heavy.
Mitigation
- Vestiga sets Ollama's
num_threadtoCPU_count - 1to leave one core free for the UI — so other apps stay responsive, but OCR itself is still slow. - If OCR is unusable on your hardware, switch to a smaller language model:
Settings → Detectors → AI model→qwen3:1.5b-instruct-*. Quality drops but runs fine on 4 cores. (Setting is planned for 0.2.)
Either the image has no text, or the glm-ocr model isn't loaded. Verification:
ollama listShould include glm-ocr. If not: ollama pull glm-ocr. Restart Vestiga.
Symptoms: OCR instantly fails, summaries don't generate, rename says "AI engine unavailable".
tasklist /FI "IMAGENAME eq ollama.exe"— should list a running process.- If missing, start it:
ollama serve(leave the terminal open) or restart Vestiga — our startup tries to spawn the daemon. - On rare occasions the installer's tray-suppression step (which kills
ollama app.exe) also leaves the daemon stopped; this is fixed by re-launching Vestiga.
Clipboard / detection#
Checklist:
Settings → Detectors— the relevant detector is enabled.- The copied text is a single line and made almost entirely of the detectable value (≤2 stray letter/digit chars outside the match). Sentences containing a detectable fragment deliberately don't fire.
Settings → Privacy & data— the external API for that detector (CoinGecko, DexScreener, open.er-api.com) isn't switched off.
The dedup setting (on by default) hides consecutive identical copies. If you re-copy the same value twice in a row, only one entry lands in history by design — toggle it off in Settings → Clipboard if you need every copy.
Smart rename#
Windows 11 Explorer tabs: the rename hotkey reads the selected file from the currently focused tab. If you have multiple Explorer windows or tabs open, click inside the tab with the file you want to rename, then press the hotkey.
Symptom of a cold LLM or a tiny/empty file. The model needs content to reason about. For very small files the best it can do is generic. Open a larger sample first to warm the model.
Uninstall#
By design. %LocalAppData%\Vestiga\ is preserved on uninstall so your clipboard history and screenshots aren't lost when re-installing. Wipe it manually if you want a clean reset:
rmdir /s "%LocalAppData%\Vestiga"Also by design — Ollama may be used by other tools on your system. If you only installed it for Vestiga, remove it from Add or Remove Programs separately.
%LocalAppData%\Vestiga\settings.json. We triage new reports within a day during beta.