tacit

Harness your tacit knowledge into AI agent context — on-device, always-on transcription.

curl -fsSL https://raw.githubusercontent.com/sangmin7648/tacit/main/install.sh | sh
tacit setup
tacit listen   # start capturing

---

Spoken ideas disappear. tacit transcribes them on-device, classifies them with Ollama, and surfaces them as live context in any AI conversation — automatically.

---

How it works

speak → capture → VAD → STT → classify → store → retrieve

1. Capture — Records microphone and system audio simultaneously in real time
2. Process — Voice Activity Detection filters silence; Whisper transcribes speech on-device
3. Classify — Ollama extracts title, category, keywords, and summary from the transcript
4. Store — Saves a structured Markdown entry to ~/.tacit/<category>/
5. Retrieve — /tacit.knowledge searches your knowledge base from inside any Claude conversation

---

Features

• Fully automatic — speak naturally; tacit handles transcription, classification, and storage without any manual steps
• On-device STT — powered by whisper.cpp; no audio ever leaves your machine
• Dual audio sources — captures microphone and system audio simultaneously
• Language-agnostic — Whisper auto-detects language; works with Korean, English, or mixed conversation
• AI-native retrieval — first-class Claude Code CLI skill integration for in-conversation search

---

Use with AI

After tacit setup, two skills are available inside Claude Code conversations:

/tacit.knowledge — search your spoken history

/tacit.knowledge summarize the search ranking discussion from earlier
/tacit.knowledge find the API design we talked about last week
/tacit.knowledge any ideas about the onboarding flow from last month?

/tacit.memorize — save the current conversation

/tacit.memorize
/tacit.memorize skill development   # optional hint guides categorization

Analyzes the current Claude conversation thread and saves it as a structured knowledge entry — automatically available to future /tacit.knowledge queries.

---

Configuration

Run tacit setup to generate two files in ~/.tacit/:

• config.yaml — a generated reference listing every field and its current default. Regenerated by tacit setup; do not edit it.
• config-override.yaml — where your own settings go. Only the fields you uncomment here override the defaults.

tacit config view prints the merged result, tagging each field [default] or [override]. tacit config edit opens the override file.

Transcription

Field	Type	Default	Description
whisper_model	string	large-v3-turbo	Whisper model name, resolved as ggml-<name>.bin against the whisper.cpp model repo and downloaded on first use — e.g. base, small, medium, large-v3, large-v3-turbo. Larger = more accurate, slower.
language	string	auto	Transcription language code (auto, en, ko, …). Pinning it to the language you actually speak markedly reduces wrong-language transcriptions.
initial_prompt	string	""	Vocabulary hint biasing decoding — useful for names, jargon, and acronyms Whisper keeps getting wrong.
experimental	bool	false	Enables non-speech token suppression during decoding plus ~192ms of VAD pre-roll padding, so a late-firing VAD onset doesn't clip the first word.
transcript_denylist	list	[]	Extra phrases to strip from transcripts, added to a built-in list of stock sentences Whisper hallucinates over silence (video outros and the like). A sentence is dropped only when a listed phrase makes up most of it; matching ignores case, spacing and punctuation.
dedup_window	duration	3h	Drops a transcript whose text has already been stored several times within this rolling window — the fingerprint of a stock hallucination, which recurs verbatim far more than real speech does. The first two occurrences per window are always kept, so a genuinely repeated remark survives. 0 disables.
min_char_rate	float	0.2	Drops a live transcript carrying too few characters for the length of audio it came from (letters and digits per second) — what is left when Whisper transcribes a stock phrase over a stretch it otherwise read as silence. Set low, so only unambiguous cases are caught. 0 disables.

Segmentation

Field	Type	Default	Description
min_speech_duration	duration	5s	Minimum segment length to process. Shorter segments are discarded.
silence_duration	duration	3s	Silence required to end a speech segment.
max_segment_duration	duration	30s	Caps a single segment sent to Whisper. Longer speech is split and transcribed as it goes, keeping memory bounded. 0 disables the cap.
max_session_duration	duration	5m	Caps how long transcribed text accumulates before it is classified. Continuous speech never triggers the silence-based flush, so this keeps a long meeting from becoming one giant entry. With max_segment_duration at 0 this doubles as the split boundary, since text only accumulates at a split. 0 disables the cap.
speech_threshold	float	0.5	VAD confidence threshold (0–1). Higher = more conservative.
energy_threshold	int	200	Audio energy gate. Frames below this value are rejected before VAD.

The three segmentation durations can be set per source, since a microphone and a meeting on your speakers behave differently. A source-specific value wins over the shared one; leave it at 0 to inherit.

Suffix	mic_ default	speaker_ default
min_speech_duration	2s	5s
silence_duration	10s	3s
max_segment_duration	30s	30s

Capture and classification

Field	Type	Default	Description
capture_mic	bool	true	Capture the microphone.
capture_speaker	bool	true	Capture system audio via ScreenCaptureKit. Requires Screen Recording permission.
llm_provider	string	ollama	Classification backend: ollama or claude.
llm_model	string	qwen3.5	Model used for classification. For claude, a Claude Code CLI model name such as haiku.
skill_agent	string	claude	Agent the /tacit.* skills are installed for.

A transcript that reaches classification is always stored. If the classifier errors out, returns nothing usable, or fills in only some of the fields, the entry is repaired — a missing title is taken from its opening words, a missing category becomes unsorted — rather than being dropped, so nothing you actually said goes missing from tacit list.

The one exception is a deliberate skip: the classifier still discards a transcript it judges to carry nothing worth keeping (filler sounds, bare acknowledgements, call-connection chatter), and logs the text it dropped.

---

Architecture

graph LR
    MIC[Microphone\n16kHz mono] --> CAP[Capture\nmalgo]
    SYS[System Audio\nScreenCaptureKit] --> CAP
    CAP --> VAD[VAD\nten-vad / Silero]
    VAD --> BUF[Segment Buffer]
    BUF --> STT[STT\nwhisper.cpp]
    STT --> CLS[Classify\nOllama]
    CLS --> KB[Knowledge Base\n~/.tacit/]

Loading

Storage format — each entry is a plain Markdown file:

---
title: "Search ranking discussion"
category: "dev"
created_at: "2026-04-14T15:30:45+09:00"
keywords: ["search", "ranking", "BM25", "lexical", "recall"]
---

One-sentence AI-generated summary.

---

Raw transcribed text from speech.

Entries are stored under ~/.tacit/<category>/YYYYMMDD-HHMMSS.md — plain files, no proprietary database, fully editable.

---

Requirements

• macOS (Apple Silicon)
• Claude Code CLI

---

Build from source

Requirements: Go 1.23+, CMake, macOS

git clone --recursive https://github.com/sangmin7648/tacit.git
cd tacit
make build
make install   # installs to ~/.local/bin/tacit

If ~/.local/bin is not in your PATH:

export PATH="$HOME/.local/bin:$PATH"

Run the end-to-end test to verify the full pipeline:

make e2e-test

Contributing

Issues and pull requests are welcome. Please open an issue first for significant changes.

make test       # run unit tests
make e2e-test   # build + process test audio through full pipeline

Note: Do not run go build ./... directly — pkg/stt uses CGo against whisper.cpp and requires make build to compile first.