> ## Documentation Index
> Fetch the complete documentation index at: https://docs.augent.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Changelog

> All notable changes to Augent.

Format follows [Keep a Changelog](https://keepachangelog.com/).

***

## \[2026.3.27] - 2026-03-27

### Added

* **`visual` MCP tool:** extract visual context from video at moments where audio alone isn't enough. Four modes: query (semantic search for specific moments), auto (autonomous detection of UI actions, demonstrations, spatial references), manual (user-specified timestamps), and assist (flags visual gaps for manual screenshots, ideal for talking-head videos). Frames are saved directly into the Obsidian vault with `![[filename.png]]` wikilink embeds.
* **`take_notes` visual integration:** pass `visual="the workflow setup"` to `take_notes` for one-shot notes plus screenshots in a single call.
* **Smart frame selection:** extracts three candidate frames per timestamp and picks the one with highest edge density, capturing the sharpest UI or screen content.
* **Duplicate frame detection:** perceptual hashing drops near-identical frames automatically.
* **Intro B-roll suppression:** first 30 seconds of a video require much higher confidence to prevent stock footage false positives.
* **Obsidian vault auto-detection:** reads vault path from Obsidian's own config. No hardcoded paths.

***

## \[2026.3.24] - 2026-03-24

### Added

* **Obsidian graph view integration:** YAML frontmatter on all memory `.md` files, `[[wikilinks]]` between related transcriptions, and MOC hub files for tag clusters.
* **`rebuild_graph` MCP tool:** migrate files, compute related links, and generate MOCs in one command.
* **Chronological neighbor links:** every new transcription links to the previous one, preventing orphan nodes.

### Changed

* **`take_notes` output switched from `.txt` to `.md`:** notes files now include auto-generated YAML frontmatter with tags, source links, and style metadata.
* **Tag changes sync to `.md` frontmatter:** `add_tags` and `remove_tags` update the file in real time.
* **Multi-word tags hyphenated:** prevents Obsidian from splitting them into separate tags.

***

## \[2026.3.21] - 2026-03-21

### Added

* **User configuration (`~/.augent/config.yaml`):** set defaults for model\_size, output\_dir, notes\_output\_dir, clip\_padding, context\_words, tts\_voice, and tts\_speed. Per-call arguments always override config values. Falls back to JSON if PyYAML is not installed. No config file required — all values have sensible defaults.
* **`disabled_tools` config key:** list tool names to hide from MCP clients.

***

## \[2026.3.20] - 2026-03-20

### Added

* **`tag` MCP tool:** add, remove, or list tags on any transcription for organizing and filtering memories by topic.
* **Semantic auto-tagging:** new transcriptions are automatically tagged using sentence-transformer embeddings: no LLM required. Works on both Web UI and MCP flows.
* **Web UI tag filtering:** clickable tag pills above the Memory list with inline counts. Memory cards display up to 5 tag pills each.
* **Tags API:** `GET /api/memory/tags` returns all tags with usage counts.

### Changed

* **Clip export improvements:** highlights and deep\_search clips now use full time ranges instead of symmetric padding around a single point. Default padding increased from 5s to 10s.
* **Tagging flow:** `transcribe_audio` and `take_notes` responses include the existing tag library so Claude reuses categories instead of creating duplicates.

### Removed

* **Regex auto-tagger:** replaced by semantic tagging.

***

## \[2026.3.19] - 2026-03-19

### Added

* **`highlights` MCP tool:** export MP4 clips of specific moments, auto-pick the best or target exactly what you want. Two modes: auto (AI picks top moments by content density) or focused (find moments matching a specific topic, person, or concept). Supports optional clip export with configurable padding.
* **Auto-tagging:** transcriptions created via `take_notes` now automatically extract and store tags for faster discovery.
* **`search_audio` and `deep_search` clip export:** both tools now support `clip` and `clip_padding` parameters to export MP4 clips around keyword matches.
* **Expanded multilingual support:** multilingual section updated with 99 supported languages.

### Changed

* **Tool descriptions:** "extract" → "export" across highlights tool descriptions for consistency. Em dashes replaced with commas in tool table descriptions.
* **CI dependency bumps:** `actions/labeler` v5 → v6, `actions/stale` v9 → v10, `actions/checkout` v4 → v6, `actions/setup-python` v5 → v6, `huggingface-hub` upper bound updated.

### Fixed

* **CodeQL security fix:** validated clip reveal path against tracked clips to prevent path traversal.
* **`take_notes` from memory:** new `output_path` parameter allows saving notes from memory transcripts without requiring a URL.
* **Test coverage:** added tests for highlights tool and auto-tagging.
* **Formatting:** black formatting applied to all modified files, ruff lint fixes for unused variables.

***

## \[2026.3.15] - 2026-03-15

### Added

* **Web UI visual clip export:** click the film icon on any search result to create a region on the waveform, or drag-select any range manually. A clip toolbar appears below the waveform with ±1s/±5s nudge buttons for precise boundary adjustment, a preview button to hear exactly what will be exported, and an Export MP4 button. Keyboard shortcuts: `Space` to preview, `Enter` to export, `Esc` to close.
* **WaveSurfer Regions plugin:** interactive regions on the audio waveform with draggable edges and body repositioning. Only one clip region active at a time.
* **Expanded Web UI tips:** 8 tips (up from 3) covering clip export workflow, Memory tab re-search, parallel tabs, YouTube timestamps, and cross-memory search.

***

## \[2026.3.12] - 2026-03-12

### Added

* **`clip_export` MCP tool:** export a video clip from any URL for a specific time range. Downloads only the requested segment: not the full video. Supports YouTube, Instagram, TikTok, Twitter/X, and 1000+ sites via yt-dlp.
* **Multilingual translation flow:** when `transcribe_audio` or `take_notes` detects non-English audio, a translation offer is returned. Accepting stores a clean English `(eng)` sibling file in memory alongside the original transcription.
* **Web UI "Show Transcript" button:** new document icon on memory cards reveals the `.md` transcript file in Finder, separate from the audio file reveal.
* **Web UI "Show Audio" / "Show Transcript" in detail view:** two distinct buttons to reveal either the audio source or the transcript file.
* **Web UI re-search from memory:** search a previously transcribed file again without re-uploading.

### Changed

* **Download filenames now include video ID:** `%(title)s [%(id)s].%(ext)s` prevents overwrites when multiple videos share the same title (e.g., multiple Instagram reels from the same account).
* **Download filenames sanitized to ASCII:** `--restrict-filenames` replaces Unicode characters (smart quotes, accented characters, etc.) with safe ASCII equivalents, preventing broken file paths in downstream tools.
* **Translation `(eng)` files are now clean English text:** no timestamps, no per-segment mapping: just the full translated text with metadata header.
* **Translation offer moved to `transcribe_audio` level:** works for any transcription call, not just `take_notes`. Removed fragile global state.
* **Clip export now overwrites existing files:** `--force-overwrites` ensures re-exporting a clip with different padding replaces the previous file instead of silently skipping.
* **Default clip padding increased to 15 seconds:** up from 10s (MCP) and 5s (CLI) for better context around keyword matches.
* **Web UI default port changed to 8282:** moved from 9797 to avoid the crowded 9000s range.

### Fixed

* **CI workflow auth failures:** downgraded `actions/checkout@v6` → `@v4` and `actions/setup-python@v6` → `@v5`; added explicit `permissions: contents: read` to Tests workflow.
* **Test count mismatch:** updated expected tool count from 16 → 17 after `clip_export` was added.
* **Ruff lint errors:** removed f-strings without placeholders, unused variables, `dict()` calls.
* **Black formatting:** reformatted all modified files to pass CI.
* **Handler test coverage:** added 59 tests for `download_audio`, `clip_export`, `transcribe_audio`, `chapters`, `batch_search`, and `separate_audio`: covering command construction, flag verification, parameter defaults, and error handling.

***

## \[2026.3.11] - 2026-03-11

### Added

* **Web UI waveform for URL downloads:** downloading audio from a URL now renders an interactive waveform with full playback controls (play/pause, skip to start/end, volume slider) — same experience as file uploads
* **Web UI download spinner:** animated wormhole spinner replaces the progress bar during audio downloads for a cleaner look
* **Web UI transcription progress bar:** real-time progress bar shown in the results panel during transcription
* **Web UI Show in Finder button:** folder icon on memory cards to reveal the audio file (or `.md` fallback) in Finder
* **Web UI URL-based memory caching:** pasting the same URL again skips re-downloading: instant cache hit by source URL + model size

### Changed

* **Web UI memory stats:** transcription count and placeholder text are now larger and use the green accent color for better readability
* **Web UI keyword placeholder:** updated to "wormhole, open source, workflow"
* **Web UI progress/spinner location:** moved from the log box to the results panel where it's more visible

### Fixed

* **`augent-web` startup:** server no longer silently exits on launch; prints a clean "WebUI is live at" message
* **`_kill_port` killing own process:** added PID check so the server doesn't terminate itself on startup
* **Memory not saving from Web UI downloads:** macOS `tempfile.mkdtemp()` created dirs in `/var/folders/` which failed the `/tmp` security check — now forces `/tmp` as the base directory
* **Memory card title overlapping trash icon:** added padding so long titles no longer clip behind the delete button
* **Memory list clipping at bottom:** added bottom padding so the last card isn't cut off by the panel edge
* **Audio endpoint 404 on macOS:** resolved `/tmp` → `/private/tmp` symlink issue and URL-encoded audio paths
* **CI build failure:** fixed black formatting issues

***

## \[2026.3.10] - 2026-03-10

### Added

* **Persistent source URLs:** Source URLs from any platform (YouTube, Twitter/X, TikTok, Instagram, SoundCloud, and 1000+ sites) are now stored permanently by audio file hash in a dedicated `source_urls` table. Any future search or transcription of the same file — even after restarts, from a different path — automatically links back to the original source. No manual URL entry needed.
* **Web UI Memory Explorer:** browse, search, view, and delete stored transcriptions from the browser. Each entry shows title, duration, model size, and date. YouTube-sourced entries are marked with a YT badge.
* **Web UI memory deletion:** trash icon on each memory card with confirmation dialog. Removes the transcription, embeddings, and `.md` file.
* **Web UI YouTube timestamp linking:** after uploading a file, paste a YouTube URL inline in the results to retroactively add clickable timestamp links to all matches.
* **Web UI URL paste:** paste a YouTube or video URL directly in the search view to download and search audio without leaving the browser.
* **Web UI Share as HTML:** download any transcript as a self-contained HTML page with all styling inline.
* **Web UI Show in Finder:** reveal the original audio file in Finder (macOS), file manager (Linux), or Explorer (Windows). Uses AppleScript on macOS for reliable handling of special characters in paths.
* **Web UI favicon:** custom Augent favicon embedded as base64, ships with every install.
* **Web UI PNG banner:** Augent logo displayed in the log area, served from the package.
* **Source URL in `.md` files:** transcription markdown files now include a `**URL:**` line in the metadata header when a source URL is available.

### Changed

* **Web UI title:** browser tab now reads "Augent Web UI" instead of "Augent"
* **Web UI tagline:** displays "Web UI v{version}" dynamically from package version
* **Web UI keywords:** changed from single-line input to resizable textarea
* **Web UI results:** row hover animation changed from border-left (caused text jitter) to clean `translateY(-1px)` on tbody rows only
* **Web UI snippets:** markdown `**` bold markers stripped from keyword matches (the UI already bolds keywords with CSS)
* **File upload naming:** uploaded files now preserve their original filename in memory instead of storing as temp file names

### Fixed

* **YouTube timestamp linking in Web UI:** `lastGrouped` was not being set in the file upload SSE path, causing the "Link timestamps" button to silently fail
* **Results header hover:** `Time | Context` header row no longer animates on hover (scoped to `tbody tr` only)
* **Show in Finder reliability:** switched from `open -R` to AppleScript on macOS for paths with special characters

***

## \[2026.3.8] - 2026-03-08

### Added

* **`separate_audio` MCP tool:** audio source separation using Meta's Demucs v4. Isolates vocals from music, background noise, and other sounds. Feeds directly into the transcription pipeline for clean results on noisy recordings.
* **`separator` optional dependency group:** `pip install augent[separator]` installs Demucs. Also included in `augent[all]`.
* **Hash-based separation caching:** separated stems are stored at `~/.augent/separated/` by file hash. First run processes, every run after is instant.
* **Vocals-only mode:** default two-stem separation (vocals + no\_vocals) is faster than full 4-stem. Set `vocals_only: false` for all 4 stems (vocals, drums, bass, other).
* **Installer support for Demucs:** `install.sh` now includes separator in the extras fallback chain and verifies demucs during package verification.
* **78 new tests:** CLI, TTS, speakers, and clips modules now fully tested (133 to 211 total)
* **CodeQL security scanning:** weekly + on every push/PR
* **Dependabot:** daily automated dependency and GitHub Actions updates
* **Makefile:** `make test`, `make lint`, `make fmt`, `make check` for developer convenience
* **Discord badge and link** in README

### Changed

* **Speaker diarization upgraded to pyannote-audio:** replaced the abandoned simple-diarizer (155 stars, last release 2022) with pyannote-audio (9.3k stars, actively maintained). Same tool interface, same caching, dramatically better accuracy. Requires a free Hugging Face token.

### Improved

* **CI pipeline:** added ruff + black enforcement, pip caching, Python 3.13 to test matrix
* **Ruff config:** migrated to `[tool.ruff.lint]` namespace, added per-file ignores for intentional patterns
* **Coverage reporting:** pytest-cov on Python 3.12 in CI with `[tool.coverage]` config

### Fixed

* **Insecure temp file** in `mcp.py`: replaced `tempfile.mktemp` with `NamedTemporaryFile`
* **Path validation** in `memory.py`: `realpath` + `isfile` check before file access
* **Lint violations:** 66 issues resolved across all modules (unused imports, raise-from, set comprehensions)
* **Formatting:** entire codebase now passes black

***

## \[2026.2.28] - 2026-02-28

### Added

* **`context_words` parameter on `deep_search` and `search_memory`:** control how much context each result returns. Default 25 words (unchanged). Set to 150 for full evidence blocks when Claude needs to answer a question, not just find a moment
* **`dedup_seconds` parameter on `deep_search` and `search_memory`:** merge matches from the same time range to avoid redundant results. Default 0 (off). Set to 60 for Q\&A workflows
* **File output on all search and transcription tools:** `transcribe_audio`, `search_audio`, `deep_search`, `search_proximity` now accept an `output` parameter for saving results directly to disk
* **XLSX export:** `.xlsx` for styled spreadsheets with bold headers and formatted timestamps, `.csv` for plain data. Auto-detected from file extension
* **Per-segment timestamps on `transcribe_audio`:** responses now include a `segments` array with `start`, `end`, `timestamp`, and `text` per segment instead of one flat text string
* **Audio trimming on `transcribe_audio`:** `start` and `duration` parameters (in seconds) to transcribe specific sections without manual ffmpeg

### Improved

* **Consistent `output` parameter:** all search and transcription tools now follow the same export pattern `search_memory` introduced in 2026.2.26

***

## \[2026.2.26] - 2026-02-26

### Added

* **`search_memory` tool:** search across ALL stored transcriptions in one query, no audio\_path needed
* **Keyword and semantic modes:** `search_memory` defaults to literal keyword matching; opt into meaning-based search with `mode: "semantic"`
* **CSV export:** optional `output` parameter on `search_memory` saves results as a CSV file
* **25-word snippets:** all search tools now return consistent \~25-word context snippets with keyword highlighting

### Improved

* **Keyword highlighting:** matched keywords shown in **bold** across all search results (search\_audio, deep\_search, search\_memory, search\_proximity)
* **CLI:** `augent memory search "query"` with `--semantic` and `--top-k` flags

***

## \[2026.2.21] - 2026-02-21

### Changed

* **Cache rebranded to Memory:** all `cache` commands, paths, and APIs renamed to `memory` (`~/.augent/cache/` → `~/.augent/memory/`)

### Improved

* **Installer UX:** animated spinners, paced output, and race condition fix for `curl|bash` piped installs
* **ASCII banner** for CLI and installer using pyfiglet

***

## \[2026.2.16] - 2026-02-16

### Added

* **OpenClaw integration:** skill package for ClawHub + `augent setup openclaw` one-liner
* **Installer auto-detects OpenClaw** and configures MCP alongside Claude
* **MCP protocol tests:** 33 tests covering routing, tool listing, and error handling

***

## \[2026.2.15] - 2026-02-15

### Fixed

* **TTS no longer blocks MCP:** runs in background subprocess with job polling
* Installer correctly selects framework Python for MCP config on macOS

***

## \[2026.2.14] - 2026-02-14

### Improved

* **Quiz checkbox syntax:** answer options render as Obsidian checkboxes
* Answer key formatting enforced (bold number + letter, em dash, explanation)
* Claude always routes video URLs through `take_notes`, never WebFetch

***

## \[2026.2.13] - 2026-02-13

### Added

* **`save_content` mode for `take_notes`:** bypasses Write tool, ensures post-processing runs
* Installer auto-installs Python 3.12 when only 3.13 is available

### Fixed

* Installer eliminates silent failures and verifies all packages
* `take_notes` embeds absolute file paths in Claude instructions
* Skip re-downloading dependencies on reinstall

***

## \[2026.2.12] - 2026-02-12

### Improved

* **Lazy imports** for optional dependencies: installing mid-session works without restart
* Preserve WAV file when ffmpeg conversion fails in TTS

***

## \[2026.2.9] - 2026-02-09

### Added

* **Text-to-speech:** Kokoro TTS with 54 voices across 9 languages
* **`read_aloud` option** for `take_notes`: generates spoken MP3 and embeds in Obsidian

***

## \[2026.2.8] - 2026-02-08

### Added

* **`identify_speakers`:** speaker diarization using pyannote-audio
* **`deep_search`:** semantic search using sentence-transformers (find by meaning, not keywords)
* **`chapters`:** auto-detect topic boundaries with embedding similarity
* **5 note styles** for `take_notes`: tldr, notes, highlight, eye-candy, quiz
* **Obsidian .txt integration guide:** full setup for live-synced notes

### Changed

* Renamed `list_audio_files` → `list_files`, defaults to all common media formats
* Enforced `tiny` as default model across all tool schemas

***

## \[2026.2.7] - 2026-02-07

### Added

* **`take_notes` tool:** one-click URL to formatted notes pipeline (download + transcribe + save .txt)

***

## \[2026.1.31] - 2026-01-31

### Added

* **Title-based cache lookups:** search cached transcriptions by name
* **Markdown transcription files:** each cached transcription also saved as readable `.md`

***

## \[2026.1.29] - 2026-01-29

### Changed

* **Python 3.10+ required** (dropped 3.9 support)

### Fixed

* Homebrew Python compatibility (PATH, PEP 668, absolute paths)
* Pinned yt-dlp to stable version for reliable downloads
* Installer handles Homebrew permission issues gracefully

***

## \[2026.1.26] - 2026-01-26

### Added

* **`audio-downloader` CLI tool:** speed-optimized with aria2c (16 parallel connections)
* **`download_audio` MCP tool:** Claude can download audio directly
* Model size warnings for medium/large (resource-intensive)

### Fixed

* aria2c downloader-args format causing download failures
* Web UI default port changed from 8888 to 8282

***

## \[2026.1.24] - 2026-01-24

### Added

* **Web UI v1:** local web interface with failproof startup
* **CI/CD:** GitHub Actions testing on Python 3.10, 3.11, 3.12
* **Professional installer:** one-liner `curl | bash` setup
* Logo and branding

***

## \[2026.1.23] - 2026-01-23

### Added

* **Initial release**
* **MCP server** exposing tools for Claude Code and Claude Desktop
* **Transcription engine** powered by faster-whisper with word-level timestamps
* **Keyword search** with timestamped matches and context snippets
* **Proximity search:** find where keywords appear near each other
* **Batch processing:** search multiple files in parallel
* **Three-layer caching:** transcriptions, embeddings, and diarization in SQLite
* **CLI** with search, transcribe, proximity, and cache management commands
* **Export formats:** JSON, CSV, SRT, VTT, Markdown
* **Cross-platform support:** macOS, Linux, Windows