What is Talkthrough?
Talkthrough, from the GitHub user korovin-aa97, is a Python MCP server that processes a video or audio file on your machine: Whisper speech-to-text, scene-change key frames, OCR of on-screen text, optional speaker labels, and timestamps mapped to real clock time. An agent then queries the stored job for transcript slices, frames and search hits, and the bundled prompts turn a recording into a bug report draft, findings list, spec, backlog or meeting actions.
It is MIT-licensed, installs from PyPI with uvx, and is listed in the official MCP registry. Version 0.4.4 was released on 23 October 2026; the repository had 34 stars and 14 open issues on 11 November. A process_url tool, added in 0.4.0, downloads one public video from a direct link, YouTube or any page yt-dlp can read.
How to install Talkthrough
uvx --python ">=3.11,<3.14" "talkthrough-mcp[diarization,url]"Is Talkthrough worth it?
Talkthrough's useful idea is wall-clock anchoring. Every timestamped result carries both the position in the video and an ISO 8601 real time, resolved from a recorded_at override, QuickTime or container metadata, or the file's modified time, with a confidence label for each. That turns "the spinner froze here" into a 30-second window to grep in server logs, which a multimodal model watching the clip cannot give you. The correlate-with-logs prompt is built around it.
The server exposes 9 tools and 6 prompts. process_media and process_url build a job, idempotent by content hash so a second call is instant; get_transcript, get_frames (up to 6 per call), get_moment, search over transcript and OCR text, extract_frame at an exact timestamp, label_speakers and list_jobs read it back. The agent pulls frames lazily rather than receiving a frame dump, so an hour-long recording does not land in context at once. The FAQ's own comparison is fair: for a short clip that is not sensitive, uploading it to a multimodal model is simpler.
The catch is weight. A cold run needs about 2 GB of memory, the default small Whisper model is a 464 MB download, ffmpeg falls back to an 80 MB bundled build, and the docs warn a cold uvx environment can exceed a client's start-up timeout. Processing runs at roughly 3 times real time on an Apple Silicon CPU, so a 2-minute clip takes about 40 seconds. Keyframes are scene changes, not motion analysis, and a glitch between them can be missed until extract_frame checks the instant.
On security the README is unusually precise. Recordings never leave the machine, there is no telemetry, and diarization keeps no voiceprints on disk. process_media accepts any local path the agent names and runs ffmpeg on it with your privileges. process_url is the one runtime network tool; it re-checks every redirect against private and cloud-metadata addresses for direct links and YouTube, but the security policy admits that on other video pages yt-dlp's own requests are not re-checked, and YouTube's JavaScript challenge runs locally in Deno. Transcript and OCR text from someone else's recording can carry prompt injection to your agent, which the policy calls inherent to the domain.
Where it earns its place
A support engineer at a SaaS company receiving customer screencasts
Turn a customer's narrated .mov into a GitHub issue with the exact frame and time
Run the bug prompt on the file; the agent quotes the remark, attaches the frame and gives a wall-clock window. Filing it uses your own GitHub tooling, since Talkthrough never posts anything. Free; budget a minute or two of local processing per short clip.
An on-call SRE reviewing a recorded incident call
Match what people said during the outage to entries in the server logs
Pass recorded_at if the recorder's metadata is weak, then use correlate-with-logs to map each remark to a time window and grep the logs for it. Speaker labels need diarize=true and about 1 to 1.5 GB extra memory on long calls.
A product manager after a recorded design workshop
Get a spec with quoted decisions and open questions
Use the spec-from-workshop prompt; quotes stay verbatim with timestamps so each decision can be checked against the recording. Pass num_speakers when the headcount is known.
Before you commit
- Setup
- Install uv, then add the server with uvx --python ">=3.11,<3.14" "talkthrough-mcp[diarization,url]"; README snippets cover Claude Code (server or plugin), Claude Desktop, Cursor, Codex CLI, Gemini CLI, VS Code, Zed, Windsurf and others. The first process_media call downloads the Whisper and OCR models.
- Integrations
- 9 tools (process_media, process_url, get_transcript, get_frames, get_moment, search, label_speakers, extract_frame, list_jobs) and 6 prompts. Write actions are limited to its own job store at ~/.talkthrough; a CLI shares the same store.
- Data and privacy
- Reads any local media path the agent passes and stores transcripts, frames and OCR under ~/.talkthrough; no telemetry and no upload. Network use is one-time model and ffmpeg downloads and process_url's single download, which stores a hash rather than the raw URL. Stdio only, no port. Transcript and OCR text is untrusted input to the agent.
- Cost at scale
- Free. Local cost is CPU and memory: about 3 times real time on Apple Silicon with the default model, larger Whisper models of about 1.5 GB for non-English narration, and recordings capped at 7,200 seconds by default.
- Limits
- 14 open issues. URL ingestion takes one public video at a time, no logins, playlists or live streams; overlapping voices beyond two degrade speaker labels; tiny or low-contrast text OCRs poorly; Windows has no cross-process lock.
What’s good
- Wall-clock timestamps for log correlation
- Fully local transcription, OCR and speaker labels
- Idempotent jobs answer follow-ups without reprocessing
- Six ready workflow prompts, including bug drafts
What’s not
- About 2 GB of memory for a cold run
- Large first-run model and ffmpeg downloads
- Keyframes can miss glitches between scene changes
- Third-party recordings expose the agent to injected text
Best for
Developers and QA engineers who receive narrated screen recordings and want an agent to turn them into evidence-backed issues tied to server-log times.
Talkthrough alternatives
Talkthrough: common questions
- Is Talkthrough worth it?
- A careful, well-documented way to hand an agent a screen recording instead of a written bug report, worth the 2 GB of memory and the first-run downloads. We rate it 3.8 out of 5.
- How much does Talkthrough cost?
- Free and open source (MIT). Runs locally with no account; no paid tier.
- What are the downsides of Talkthrough?
- About 2 GB of memory for a cold run; large first-run model and ffmpeg downloads; keyframes can miss glitches between scene changes; third-party recordings expose the agent to injected text.
- Who is Talkthrough best for?
- Developers and QA engineers who receive narrated screen recordings and want an agent to turn them into evidence-backed issues tied to server-log times.
- How much does Talkthrough cost for a team or in production?
- Free. Local cost is CPU and memory: about 3 times real time on Apple Silicon with the default model, larger Whisper models of about 1.5 GB for non-English narration, and recordings capped at 7,200 seconds by default.
- Is Talkthrough safe for company data?
- Reads any local media path the agent passes and stores transcripts, frames and OCR under ~/.talkthrough; no telemetry and no upload. Network use is one-time model and ffmpeg downloads and process_url's single download, which stores a hash rather than the raw URL. Stdio only, no port. Transcript and OCR text is untrusted input to the agent.
- What are the best Talkthrough alternatives?
- The closest alternatives we have reviewed are OpenEdit and agentic-obs (OBS Studio MCP).
Answers checked Nov 11, 2026.