<- all tokdocs
2026.09.27@saf3synt

Call.md is a real repo, but the "open-source desktop app" has no LICENSE file and sends every meeting's audio to VideoDB's hosted API

Coding
Watch on TikTok

View on TikTok ->

A 19.97 s vertical clip (1080x1920, 9:16, SDR) in a QuickTime/MP4 container tagged major_brand=isom, encoder=Lavf58.76.100, carrying one HEVC video stream (hvc1, Main profile, level 150, yuv420p, 30/1 fps, 599 frames, 208,311 bps) and one AAC HE-AACv2 stereo audio stream (44.1 kHz, 64,661 bps, 17.972 s), 682,512 bytes total at an overall 273,460 bps, served by TikTok as format_id=bytevc1_1080p_273456-1; posted by @saf3synt (channel saf3synt, uploader_id 7644216153246762006) at epoch 1790508634, which is 2026-09-27 11:30:34 UTC, credited to the original sound "izvirni zvok" by saf3synt, with a 366-character description that includes the repo URL https://github.com/video-db/call.md and the hashtags #AI #Opensource #CallMD #MeetingTools; at capture time (metadata.json mtime 2026-10-27 00:10:48 local) the post showed 4,924 views, 221 likes, 0 comments, 80 reposts and 265 saves. I read all ten extracted frames, frame_001 through frame_010, covering 0, 2, 4, 6, 8, 10, 12, 14, 16 and 18 seconds, against a transcript of 46 words in 3 Whisper segments. The frames are a screen recording of one web page and nothing else. Frames 001 and 002 (0 s, 2 s) show a GitHub README view behind a rotated red stamp reading "GITHUB FIND #612" and "LIVE MEETING AI", with a persistent top-left overlay badge reading "1.5k / GitHub stars"; the README screenshot underneath carries shields reading "ELECTRON 34", "LICENSE MIT", "STARS 23" and "ISSUES", the call.md wordmark, the heading "Call.md", the tagline "Turn meetings into live agent loops. Record, transcribe, and analyze meetings with real-time AI intelligence", the links "Explore the docs »" and "View Demo · Install · Report Bug", and a "Demo" heading with a collapsed player labelled call.md.mp4. Frame 003 (4 s) is a full-bleed black card showing only the call.md wordmark. Frames 004 and 005 (6 s, 8 s) cut to the live repo page for video-db/call.md, marked Public, showing "Fork 165", "Star 1.5k", branch main, "9 Branches", "0 Tags", the head commit "Merge pull request #33 from Nischay-VideoDB/..." at ba53ebe "last month", "94 Commits", an About panel linking docs.videodb.io/examples-and-tu... with the topics agentic-ai, agentic-workflow, ai, mcp, meeting, meeting-notes, note-taking, notes-app and skills, sidebar entries "1.5k stars", "5 watching", "165 forks", "Releases / No releases published", "Contributors 5", and "Languages / TypeScript 98.6% / Other 1.4%", plus a current README badge row reading "ELECTRON 42", "NODE.JS 22.12+", "REACT 19", "TYPESCRIPT 5.8", "LICENSE MIT", "STARS 1.5K", "ISSUES 1 OPEN" and "VIDEODB.IO UP". Frames 006 and 007 (10 s, 12 s) scroll to the Platform Support table, the Overview paragraph and the Features list; frame 008 (14 s) shows the Tech Stack, Prerequisites and the install line curl -fsSL https://artifacts.videodb.io/call.md/install | bash; frames 009 and 010 (16 s, 18 s) show "Building for other platforms", "MCP Server Setup" and the src/ project tree, with frame 010 adding a green call-to-action card reading "LINK IN DESCRIPTION / tap below to open it". The burned-in captions across the ten frames read, in order, "CALL MD IS", "DESKTOP APP THAT", "THROUGH YOUR MEETINGS", "IT RECORDS AND", "TRANSCRIBES CALLS AS", "RUNS AGENT LOOPS", "AFTER EACH MEETING.", "GREAT FOR TEAMS", "JUST A TRANSCRIPT." and the link card. No frame I read shows the application itself: there is no app window, no recording session, no transcript pane, no terminal and no run of the installer anywhere in the clip, and the README's embedded call.md.mp4 demo is never played. The spoken argument is three sentences long: Call.md is an open-source desktop app that runs AI through your meetings in real time, it records and transcribes live and then runs agent loops before, during and after each meeting, and it suits teams and researchers who want more than a transcript. On the hallucination check, the last Whisper segment ends at 17.56 s against an audio stream of 17.972 s and an extracted audio.mp3 of 18.024 s, so every segment closes inside the real audio and the transcript is genuine narration rather than invented text; the only Whisper damage is to the product name, which it renders as "Call, MD," where the burned-in caption reads "CALL MD IS" and the repository reads "Call.md".

The product exists, and the repo is published by the vendor whose paid API it depends on

At 0.00 s the video says "Call, MD, is an open-source desktop app that runs AI intelligence through your meetings in real-time." The name is Whisper damage. The burned-in caption in frame 001 reads "CALL MD IS" and the GitHub page in frame 004 reads video-db/call.md, so the product is Call.md.

The existence claim checks out. The GitHub API returns full_name: video-db/call.md, created 2026-02-25T12:05:19Z, default branch main, primary language TypeScript, not archived, with the description "Turn meetings into live agent loops. Record, transcribe, and analyze meetings with real-time AI intelligence", and a homepage field pointing at https://docs.videodb.io/examples-and-tutorials/ai-copilots/call-md. That docs page is live and describes Call.md as an "AI meeting assistant, live during every call."

What the video does not say is who publishes it. The repo sits in the video-db organisation. package.json lists the author as {"name": "VideoDB", "email": "support@videodb.io", "url": "https://videodb.io"} at version 1.0.4. The README closes with "Made with ❤️ by the VideoDB team". Call.md is a first-party client application for VideoDB's commercial API, not an independent community project that happens to be open source. That matters for every claim below, because the thing the app depends on is the thing its publisher sells.

The repo also carries two earlier names in its own history. The README's commit log shows 41e0a94 "init: meeting copilot" on 2026-02-26, 7cf4c5d "feat: name change to Notter" on 2026-03-24, then 5773615 "feat: name change to call.md" on 2026-03-25. The project was called Notter for roughly one day before becoming Call.md.

Correction: "open-source" is asserted by a hardcoded badge, and the repository contains no LICENSE file

At 0.00 s the video calls it "an open-source desktop app." The README badge row in frames 001, 004 and 005 shows "LICENSE MIT". Both the badge and the video are overstating what the repository actually contains.

The GitHub contents API for the repository root returns nineteen entries: .gitignore, README.md, build/, drizzle.config.ts, electron-builder.config.js, eslint.config.mjs, package-lock.json, package.json, postcss.config.js, resources/, screenshot.png, scripts/, src/, tailwind.config.ts, tasks/, tests/, tsconfig.json, tsconfig.node.json and vite.config.ts. There is no LICENSE, no LICENSE.md and no COPYING. The frame 004 file listing on screen shows the same set, with no license file visible. GitHub's license endpoint for the repo, api.github.com/repos/video-db/call.md/license, returns HTTP 404, and the repository object returns license: null.

The MIT badge is not reading a license. It is a static image. README line 440 defines it as [license-shield]: https://img.shields.io/badge/License-MIT-yellow.svg?style=for-the-badge, a fixed shields.io badge with the word MIT painted into the URL, pointing at https://opensource.org/licenses/MIT. By contrast, the stars and issues badges on the same line block are dynamic: [stars-shield]: https://img.shields.io/github/stars/video-db/call.md.svg queries GitHub live.

The only machine-readable MIT declaration in the repository is the "license": "MIT" field in package.json. That is an npm metadata field, not a grant of terms, and it is not accompanied by the license text that MIT itself requires be included in all copies. Someone forking this repo today has a vendor's stated intent and no license file to point at. The honest version of the video's first sentence is that the source is public and the publisher states MIT in package metadata, with the license text missing from the repository.

Correction: it does not transcribe locally. The audio leaves the machine and goes to VideoDB over a WebSocket

At 7.06 s the video says "It records and transcribes calls as they happen, then runs agent loops before, during, and after each meeting." Phrased next to "open-source desktop app," that invites the reading that a local app does the transcription. The README says otherwise, in its own words.

The README's "How It Works" section, under "During Recording", states: "Captures dual-channel audio (you vs them) and sends to VideoDB for real-time transcription via WebSocket". The Features list qualifies the headline capability the same way: "Dual-Channel Transcription - Separate transcription for you (mic) vs them (system audio), powered by VideoDB". The "Privacy & Storage" section states: "VideoDB Integration - Transcription and AI features require internet connectivity". The Security section is the most explicit: "The application database, settings, and logs stay on your machine. Recording, transcription, and AI inputs are sent to VideoDB when those features are used."

The language model calls are routed the same way. The Tech Stack list, visible in frame 008, reads "OpenAI SDK (6.19.0) - LLM calls via VideoDB's OpenAI-compatible API". So both halves of what the video calls "AI intelligence through your meetings" run on VideoDB's servers.

This is a genuine local-first design for storage and a hosted design for processing, and the README is straight about the split: "Local-First - Settings, meeting history, transcripts, and generated metadata are stored in the local SQLite database". The record of your meetings stays on disk. The raw audio of your meetings does not. The video compresses that into "records and transcribes calls as they happen" and loses the distinction entirely.

A VideoDB API key is a hard prerequisite, not an optional integration. The README's Prerequisites section lists "VideoDB API Key (console.videodb.io)" alongside the OS requirement, and step 2 of Getting Started is "Launch the app and enter your VideoDB API key". The docs page confirms "VideoDB API key (free tier available)". I could not verify VideoDB's pricing against a first-party source: https://www.videodb.io/pricing served a contact form with no plans, tiers or rates on it, so I am treating the cost of running Call.md at volume as unverified and citing no third-party pricing aggregator for it.

"Runs agent loops" is the repository's own marketing line, read back as a finding

At 7.06 s the video says the app "runs agent loops before, during, and after each meeting." The video's written description repeats it: "Call.md turns meetings into live agent loops."

That is the project's own tagline. README line 21 reads "Turn meetings into live agent loops. Record, transcribe, and analyze meetings with real-time AI intelligence", and the Overview at line 82 opens "Call.md turns meetings into live agent loops." The GitHub repo description field is the same sentence. The video is not independently characterising the software; it is reading the first line of the README aloud with the word order changed. No frame in the clip shows an agent loop running, because no frame shows the application at all.

What the repository actually ships under that banner is a named feature list, and it does map onto before, during and after. Before the meeting: "Meeting Setup Wizard - AI-generated probing questions based on meeting description", "Dynamic Checklist - AI creates discussion checklist from meeting context", and "Google Calendar Integration - Sync upcoming meetings". During: "Live Assist", "Conversation Metrics - Real-time monitoring of talk ratio, speaking pace (WPM), questions asked, monologue detection", "Coaching Nudges", "MCP Auto-Triggering - Detects information needs from conversation and calls your MCP tools automatically", "MCP Results Panel" and "Bookmarking". After: a three-part summary of narrative overview, key points and action items, outbound delivery to "workflow automation platforms (n8n, Zapier, CRMs)", and markdown export.

The project tree in frames 009 and 010 shows the matching services: src/main/services/copilot/ contains context-manager.service.ts, conversation-metrics.service.ts, nudge-engine.service.ts, sales-copilot.service.ts annotated "# Core orchestrator", summary-generator.service.ts and transcript-buffer.service.ts, alongside live-assist.service.ts, mcp-inference.service.ts, llm.service.ts and videodb.service.ts. The orchestrator being named sales-copilot is a useful tell about who this was built for.

Correction: "desktop app" oversells distribution. There is no release, no tag, and recording works on macOS plus one build-it-yourself platform

At 0.00 s the video says "desktop app." Frame 004 shows the GitHub sidebar contradicting the easiest reading of that on the same screen: "Releases / No releases published". The GitHub API confirms it, returning zero releases and zero tags for the repository, and the repo header in frame 004 reads "0 Tags". There is no binary to download from GitHub.

Installation is a vendor-hosted script piped into a shell. The README and frame 008 both show curl -fsSL https://artifacts.videodb.io/call.md/install | bash, served from artifacts.videodb.io rather than from the repository.

The Platform Support table, visible in frames 006 and 007, limits the capability further. macOS 12+ on Apple Silicon and Intel is "Supported" via that curl command. Windows x64 has no installer at all and is marked "Recording supported; no hosted installer yet", reachable only by building from source with npm run dist:win. Windows ARM64 shows an em dash for installer and "Recording not supported". Linux is "App features available; recording not supported", also build-from-source. The README restates it below the table: "Recording support is available on macOS arm64/x64 and Windows x64. On Linux and Windows ARM64, the full UI, MCP servers, workflows, settings, history, and markdown export remain available; starting a recording returns a clear unsupported-platform error."

So for the thing the video actually sells, recording and live transcription, there is exactly one platform you can install and run without compiling: macOS 12 or later.

The opening frames use a stale screenshot, and the star figure on screen is already out of date

The overlay badge in every frame reads "1.5k GitHub stars", and the live repo page in frames 004 and 005 agrees, showing "Star 1.5k", "1.5k stars" and "165 forks". The README screenshot behind the stamp in frames 001 and 002 does not agree with either. Its shields read "ELECTRON 34" and "STARS 23".

Both numbers date that screenshot. The stars badge is the dynamic shields.io endpoint, so a rendering showing 23 was captured when the repository had 23 stars. The Electron badge is static and versioned in the README itself: at commit 9285334 (2026-04-01) and still at 0d8eb3a (2026-08-13) it read https://img.shields.io/badge/Electron-34-..., and by the current main it reads Electron-42, changed in the 2026-08-18 pair of commits 0e0138c "docs: update security and platform guidance" and 8146d29 "build: remove capture artifact staging helper". The opening card is footage of the project from months before the post date of 2026-09-27, cut against a live page from the same week.

The live figures have moved since. Querying the GitHub API on 2026-10-27 returns 1,878 stargazers, 222 forks, 8 watchers and 8 open issues, against the 1.5k stars, 165 forks, 5 watching and "ISSUES 1 OPEN" shown on screen. Frame 004's head commit is ba53ebe, labelled "last month" on screen, and the API confirms it as the current tip of main: "Merge pull request #33 from Nischay-VideoDB/agent/fix-main-qa-findings", dated 2026-08-19T06:24:36Z. That is the repository's most recent push. Call.md had not been touched for five weeks when this video was posted, and has not been touched in the two months since.

One sidebar number does not reconcile. Frame 004 shows "Contributors 5", while the contributors API returns four accounts: omgate234 with 81 commits, Nischay-VideoDB with 8, codeAshu with 4 and lalit-videodb with 1, totalling the 94 commits shown in the header. GitHub's sidebar count can include entries the API endpoint does not return, so I am flagging the gap rather than calling either number wrong.

Unsupported: nothing in the repository describes a team or research feature

At 14.26 s the video closes with "Great for teams and researchers who want more than just a transcript." That is the only claim in the clip with no counterpart anywhere in the primary source.

Searching the full README for team, collaboration, sharing, multi-user, workspace, seat and SSO returns three hits, none of them a feature: a shared/ directory of types and schemas in the project tree, a constants path, and the "Made with ❤️ by the VideoDB team" footer. There is no shared workspace, no multi-user account model, no org or seat concept, and no cross-machine sync of meeting history. The storage model argues directly against it: "Settings, meeting history, transcripts, and generated metadata are stored in the local SQLite database", one database per machine, with sharing available only through the markdown export and the outbound webhooks to n8n, Zapier and CRMs. The word "researchers" appears nowhere in the README or the linked docs page.

One hard constraint a team would hit is also absent from the video: "2 Hour Recording Limit - Recordings stop themselves after 2 hours of active recording time, with a warning 5 minutes before; pauses and system sleep do not consume the allowance."

Key Takeaways

  • Verified: Call.md is real and open to read at github.com/video-db/call.md, created 2026-02-25, TypeScript, Electron 42, React 19, 94 commits, head ba53ebe dated 2026-08-19.
  • Correction: The video calls it "an open-source desktop app" at 0.00 s and the README shows an MIT badge. The repository contains no LICENSE file, GitHub's license endpoint returns 404, and the repo object returns license: null. The MIT badge is a hardcoded shields.io image defined at README line 440, and the only MIT declaration is the "license" field in package.json.
  • Correction: At 7.06 s the video says it "records and transcribes calls as they happen." The README states that the app "sends to VideoDB for real-time transcription via WebSocket", that "Transcription and AI features require internet connectivity", and that "Recording, transcription, and AI inputs are sent to VideoDB when those features are used." Transcription and LLM calls run on VideoDB's servers, not locally.
  • Correction: "Desktop app" at 0.00 s understates the platform limits. GitHub shows "No releases published" and zero tags, install is curl -fsSL https://artifacts.videodb.io/call.md/install | bash from the vendor's host, and the Platform Support table limits recording to macOS 12+ and a source-built Windows x64, with Linux and Windows ARM64 explicitly unable to record.
  • Partial correction: The "1.5k GitHub stars" overlay matched the repo page at post time on 2026-09-27, but the opening card behind it is stale footage showing "STARS 23" and "ELECTRON 34", a README state that ended on 2026-08-18. As of 2026-10-27 the API returns 1,878 stars and 222 forks.
  • Clarification: "Runs agent loops before, during, and after each meeting" at 7.06 s is the repository's own tagline from README line 21, not an independent assessment. The concrete features behind it are the Meeting Setup Wizard and Dynamic Checklist before, Live Assist, Conversation Metrics, Coaching Nudges and MCP Auto-Triggering during, and a three-part summary plus n8n/Zapier/CRM webhooks after.
  • Clarification: The publisher is VideoDB itself. The repo is in the video-db org, package.json names the author as VideoDB support@videodb.io, and a VideoDB API key is a listed prerequisite. This is a vendor client for a commercial API.
  • Unverified: The "great for teams and researchers" claim at 14.26 s. The README has no multi-user, shared-workspace, sync or org feature, and stores everything in a per-machine SQLite database. The word "researchers" does not appear in the README or the linked docs page.
  • Unverified: What Call.md costs to run at volume. The docs page says a VideoDB "free tier available", but videodb.io/pricing served a contact form with no plans or rates, so I am citing no pricing figure.
  • Context: No frame in the clip shows the application running. All ten frames are a GitHub page scroll plus one logo card, and the README's own call.md.mp4 demo is visible as a collapsed player but never played.
  • Context: The repo was named "meeting copilot" at init on 2026-02-26, renamed "Notter" on 2026-03-24, and renamed "call.md" on 2026-03-25. The core orchestrator file is sales-copilot.service.ts.

Resources

Published September 27, 2026. Writeup generated from a favorited TikTok.