NoCap

A desktop meeting copilot that listens to a live call and answers from the project's own memory — its docs, its git history, its saved transcripts — rather than from the model's general knowledge. It runs frameless and always-on-top on a second monitor and declares itself excluded from screen capture, so it stays off the shared screen. Speech comes through a pluggable STT chain that falls back from Deepgram to fully offline Whisper and Vosk, with speaker diarization separating your voice from the client's. Answers arrive in three styles from a single API call, each carrying an honesty label — LIVE, IN PROGRESS, BLOCKED, or WEB — where ambiguous wording is deliberately downgraded rather than rounded up, and web-sourced answers are always marked external at a single chokepoint. An over-promise guard watches what you are about to say: a local pre-filter catches claims like "that's already live", then a Claude judge compares the claim against the project's real status and offers a safer rephrase.

• Answers are grounded in the project's own docs, git history and transcripts — not the model's general knowledge
• Every answer carries an honesty label, and ambiguous wording is downgraded toward IN PROGRESS rather than rounded up to shipped
• Web-sourced answers are forced to declare themselves external at one enforced chokepoint
• Over-promise guard: a local claim filter plus a Claude judge compare what you're about to say against the project's real status
• STT degrades gracefully from Deepgram to offline Whisper and Vosk, so it still works with no key and no network
• Hand-rolled OOXML writer exports the transcript to .docx with zero runtime dependencies

An internal tool at v0.7.0 — built for my own client calls, not packaged for anyone else. Windows only: the capture-exclusion and system-audio loopback it depends on are Win32-specific.

• Electron
• JavaScript
• Anthropic Messages API
• Deepgram
• Whisper (WASM)
• Vosk
• Python
• Playwright