v0.1.3

Hardening ahead of the public announcement, from a hostile-input test pass (silent recordings, odd containers, corrupt files, offline machines).

Key fix — warm runs are now zero-network: whisper loads from the local cache first; previously huggingface_hub revalidated metadata against huggingface.co on every model load, even fully cached. Verified by running the full pipeline with all sockets blocked.

Also: tool errors name the binary instead of the full venv path, softer static-ffmpeg log, and the Quickstart now names its one real prerequisite (uv).

Verified graceful in this pass (no changes needed): videos without audio tracks, .webm/.mkv, 2-second clips, unicode filenames, corrupt files (clean exit 2), Intel-mac installs.

Full changelog: CHANGELOG.md

v0.1.2

Interop fix: all 7 tools now ship MCP ToolAnnotations (readOnly/destructive/idempotent/openWorld hints). Headless OpenAI Codex (codex exec) silently cancelled un-annotated tool calls — with hints it drives talkthrough end-to-end (verified live). Hints are honest: only process_media/extract_frame write, and only inside TALKTHROUGH_HOME. Full changelog: CHANGELOG.md

v0.1.1

Launch-day patch from a full as-a-new-user E2E pass (every install path, a real 2-minute narrated recording, contract-validated triage).

Fixed: an explicit per-call model= (tool param / CLI --model) that differs from the stored transcript's model now reprocesses instead of silently returning the old transcript. Env-default changes still reuse; force unchanged.

Docs: dead Goose one-click button removed (their deep-link endpoint 404s ecosystem-wide); troubleshooting entry for pip install on Python < 3.11.

Full changelog: CHANGELOG.md

v0.1.0 — first public release

First public release. 🎬→📋

talkthrough-mcp is a local-first MCP server that turns narrated screen recordings (or any video/audio file) into agent-ready structured data: timestamped transcript, scene keyframes, OCR'd on-screen text, and wall-clock anchoring — so "the moment I said the checkout hung" maps straight to the right window of your server logs. No LLM inside, no cloud anywhere in the path.

uvx talkthrough-mcp

Highlights

• 7 lazy retrieval tools — a 30-minute recording never floods the context window; agents pull exactly the transcript slice, moment bundle, or frame they need. Every tool description embeds 10–15 usage examples.
• 5 workflow prompts — triage-recording, spec-from-workshop, backlog-from-demo, meeting-actions, correlate-with-logs — plus a findings JSON contract.
• Wall-clock anchoring (the flagship): every timestamp carries real ISO-8601 time resolved from recording metadata, with an honest confidence ladder.
• Any of Whisper's ~99 languages, auto-detected with a confidence signal; per-call model choice (large-v3-turbo recommended for non-English). OCR script packs for non-Latin on-screen text.
• One-command install, no system dependencies: bundled ffmpeg fallback, pip-only OCR, self-downloading models. Works with Claude Code/Desktop, Cursor, Codex CLI, Gemini CLI, Cline, and any MCP stdio client — 12 engine adapters included.
• Private by design: recordings never leave your machine; the only network access is one-time model downloads. See SECURITY.md for the threat model.

Not in v1 (roadmap)

Speaker diarization (#4) · URL/YouTube ingestion (#5) · whisper.cpp backend (#6). Contributions welcome — each issue carries a design sketch.

Full changelog: CHANGELOG.md