PresenterAI

PresenterAI is a private, local-first Windows copilot for presentations, technical reviews, and coding discussions. It stays above the application you are presenting, retrieves supporting details from your local documents, and produces concise answers or copyable code without sending an entire document to the model.

PresenterAI is an unsigned development beta for Windows 11 x64. Capture exclusion and system-audio reliability still require validation on each target setup.

What it does

• Starts in Code mode with structured, copyable source-code cards; Presenter mode produces short, natural answers designed to be spoken aloud.
• Indexes PPTX, PDF, Markdown, and UTF-8 text locally with SQLite FTS5 and sends only selected evidence chunks to OpenAI.
• Captures bounded Windows system output, not microphone audio, through a restricted .NET helper. The transcript is shown as an editable draft before it can be submitted.
• Keeps the OpenAI key in Electron safeStorage (Windows DPAPI), keeps conversation state local, and sends Responses requests with store:false.
• Provides a transparent always-on-top overlay, tray controls, configurable shortcuts, fail-closed click-through recovery, and a persistent per-session USD cap.
• Makes no telemetry, analytics, account, cloud-database, or hosted-backend connection.

How it works

1. Set up — add an OpenAI API key in Settings and optionally import presentation documents.
2. Ask — type a question, or toggle system-audio capture and review the resulting transcript.
3. Generate — choose Code or Presenter and submit with Ctrl+Enter; PresenterAI retrieves up to five relevant local chunks.
4. Review — read the answer, copy code, and inspect the evidence or warning before speaking.

Run locally

Requirements

• Windows 11 x64
• Node.js 22 or newer
• .NET 8 SDK for the system-audio helper
• An OpenAI API key for transcription or generated answers

Install and start

git clone https://github.com/KanuTomer/Presentation-Helper.git
cd Presentation-Helper
npm ci
npm run helper:build
npm run dev

On first run:

1. Open Settings, save the API key, and use the connection test.
2. Open Documents to import non-sensitive PPTX, PDF, Markdown, or text fixtures.
3. Review and accept the listening disclosure before the first system-audio capture.
4. Confirm that listening is off and test Ctrl+Shift+I before enabling click-through.
5. Type a question in Copilot and press Ctrl+Enter.

The typed/document workflow still runs if the helper is unavailable; only system-audio controls are disabled.

Shortcuts

Action	Default
Focus the composer	Ctrl+Space
Hide or show PresenterAI	Ctrl+Shift+H
Start or stop system-audio capture	Ctrl+Shift+Space
Restore interaction after click-through	Ctrl+Shift+I
Submit the reviewed question	Ctrl+Enter
Cancel the active operation	Esc

Ask, hide/show, and listening shortcuts can be changed in Settings. Ctrl+Shift+I is fixed so click-through always has a known recovery path; Tray → Show PresenterAI is the second recovery method.

Build the Windows installer

npm run package:win

The command builds the helper and Electron application, creates the unsigned NSIS installer under release/, and writes SHA256SUMS.txt. Windows SmartScreen, Smart App Control, or an organization’s App Control policy may block unsigned binaries. Do not disable Windows security to run PresenterAI; use a trusted development environment or a future signed build.

Development checks

The standard offline gate does not call OpenAI:

npm audit --audit-level=high
npm run verify
npm run eval:m4
npm run eval:m6:preflight
npm run eval:m7
npm run test:e2e

Native and packaged checks:

npm run test:helper-smoke
npm run package:win
npm run test:packaged-fts
npm run test:packaged-helper
npm run test:code-integrity-environment

Live evaluation scripts are opt-in, billable, and excluded from ordinary development and CI.

Privacy and limitations

• Source documents and the FTS index stay on the computer. Only the current question, bounded local context, and selected evidence chunks are sent for an answer.
• System audio is captured in bounded helper memory. A final temporary WAV is deleted after transcription, and the transcript remains an editable in-memory draft until submission.
• The local USD cap limits PresenterAI requests, not the OpenAI account itself. Provider billing remains authoritative.
• DPAPI primarily protects the stored key from other Windows users; it cannot isolate it from every process already running as the same user.
• The transcription endpoint and Responses API have different retention rules. Review the current OpenAI API data controls before using sensitive material.
• Image-only/scanned files, visual chart interpretation, OCR, Chrome-process audio isolation, microphone capture, continuous listening, and automatic screenshots are not supported.
• setContentProtection(true) is requested on Windows, but that is not a universal security guarantee. Record every capture application/version in the compatibility matrix.

Validation status

Area	Current standing
M0–M2: Windows shell and capture protection	Implemented; fullscreen, multi-monitor, Meet, OBS, and capture-path matrices remain unsigned.
M3: typed presenter answers	The accepted Luna/Terra report remains historical evidence. The natural spoken-delivery prompt is offline-green and awaits a separately authorized live revalidation.
M4: document retrieval	Accepted offline at 50/50 top-five recall with packaged Electron FTS5.
M5–M6: capture and transcription	Source and automated checks are green; physical device/Meet and billable live acceptance remain pending.
M7: context and grounding	Source complete and offline green at 50/50; formal live acceptance remains pending.
M8: privacy, settings, and packaging	Source complete/offline green; formal acceptance depends on M2–M7.
M9: continuous listening	Not implemented and remains experimental.

See the plan alignment record, beta.4 clear-glass record, and closed technical-preview guide for the exact acceptance boundaries.

Safety boundary

PresenterAI is intended for presentations and reviews where its use is permitted and disclosed. It does not include stealth activation, proctoring or examination evasion, hidden collaboration, process-tree browser capture, or claims of being “undetectable.”