Call.md Meeting Copilot
Turn recorded meetings into live assistance, searchable transcripts, summaries, action items, and automated follow-up workflows.
The evidence clearly identifies microphone and screen-recording permissions, local storage, audio/video and prompts sent to VideoDB, loopback binding, token checks, file modes, log redaction, and webhook SSRF controls. This supports reasonably strong privilege, data-flow, and external-effect documentation. Deductions apply because MCP tools can be triggered automatically and post-meeting webhooks can send data automatically, with no evidence of per-action confirmation or fine-grained tool permissions. Sensitive-data handling has focused tests, but those tests explicitly permit plaintext secret storage when the OS keyring is unavailable, conflicting with the README's unconditional encryption-at-rest claim. Dependencies use broad ranges, and no lockfile, audit, provenance verification, or vulnerability-response evidence is supplied. In-place migration is mentioned, but backup, undo, and recovery procedures are not. VideoDB team attribution, email, and issue channels are present, while publisher identity remains externally unverified.
The README, package manifest, and tests are broadly consistent about requirements, scripts, the two-hour recording limit, and language fallback. Timer lifecycle, pause/resume, and secure-store edge cases receive targeted tests. Deductions reflect the contradiction between the security prose and plaintext fallback, plus the narrow test sample, which cannot establish end-to-end reliability for capture, transcription, models, MCP, or webhooks. Dependency, network, and platform prerequisites are clear, but the core recorder binaries are macOS-only. Troubleshooting and some failure behavior are concrete, although complete user-facing errors, retries, and recovery for VideoDB, MCP, summarization, and webhook failures are not shown.
The material thoroughly covers pre-meeting, live-meeting, and post-meeting scenarios, with separate user and developer paths and options for transcription language, stdio/HTTP MCP, workflows, and exports. Platform and capability boundaries are unusually explicit, particularly that Windows and Linux builds cannot record. Deductions apply because the MCP information-need detector and automatic trigger mechanism have no documented rules, thresholds, allowlists, false-positive handling, or per-tool controls. Language support also depends on an unspecified backend capability set. Environment-fit documentation is strong, but the functional recording product remains limited to macOS 12+.
The README has strong navigation and organization across features, project structure, installation, scripts, permissions, storage, security, and troubleshooting. Platform limitations are documented repeatedly and concretely. Deductions apply because examples focus mainly on installation and builds, without complete MCP or webhook configurations and without a broad FAQ. Naming varies slightly among call.md, call-md, and the call-md data directory. MIT is declared in the badge and package manifest, but no LICENSE text is included in the supplied evidence and the input metadata says unknown. A package version and issue references exist, but no changelog, release policy, or compatibility commitment is shown. The team and support paths are named, while individual ownership, maintenance cadence, and a dedicated security-reporting route are unclear.
Dual-channel live transcription, inline MCP results, metrics, three-part summaries, action items, full Markdown export, and history form a practical output pipeline. Live coaching and automatic tool invocation plausibly add value beyond basic meeting recording, but the evidence contains no quality comparison, outcome study, or user-result validation. Cost-benefit receives a substantial deduction because the source does not disclose transcription, model, VideoDB storage, or workflow pricing, usage consumption, latency, or resource costs; it also requires connectivity and broad audiovisual capture permissions.
Several important claims trace to package scripts, dependency declarations, and tests for language handling, recording limits, and secret storage. The README and package manifest also corroborate the stack, authorship declaration, and license declaration. Deductions apply because most feature and security claims appear only in the README, without the corresponding implementation files, configuration, lockfile, or tests—for example webhook validation, CORS, sandboxing, automatic MCP behavior, summarization, and log redaction. The access-token hashing test mirrors rather than imports the actual database implementation. The documentation usually separates current support from future plans, but some absolute security language fails to distinguish the tested plaintext-degradation case.
- MCP tools and post-meeting webhooks can create automatic external effects; verify confirmation controls, tool allowlists, argument previews, and audit logs before connecting sensitive systems.
- The tests show that secrets fall back to plaintext when the operating-system keyring is unavailable, contradicting the README's unconditional encryption-at-rest wording.
- Audio, video, and prompts are sent to VideoDB. Confirm participant consent, retention rules, server-side storage location, and applicable privacy requirements before recording.
- The one-line curl pipeline executes a remote installer; no checksum, signature, or installer contents are supplied for this static review.
- No lockfile, dependency audit, or vulnerability-response evidence is provided, and the core recorder dependency supplies macOS binaries only.
- This assessment did not execute the application or tests and did not independently verify the README's security or functionality claims.
What does this agent do, and when should you use it?
Call.md is an Electron 34 desktop meeting assistant built with React 19 and TypeScript 5.8, with full recording currently limited to macOS 12 or later. It captures the microphone, system audio, and screen, then streams separated “you” and “them” audio to VideoDB over WebSocket for real-time transcription. During a call, it measures talk ratio, speaking pace, questions, and monologues, offers contextual coaching, and can invoke connected MCP tools when the conversation creates an information need. Afterward, it produces a narrative overview, participant-attributed topic notes, and action items, with a Markdown export containing the transcript and metrics or delivery through webhooks to n8n, Zapier, and CRMs. Records are stored locally in SQLite, while transcription and AI processing require internet access and VideoDB; each recording is capped at two hours.
When recording starts, the VideoDB SDK and @videodb/recorder capture the screen, microphone, and system audio, then send speaker-separated audio to VideoDB for live transcription. conversation-metrics.service.ts calculates talk ratio, WPM, question counts, and monologue behavior; nudge-engine.service.ts issues rate-limited coaching prompts; and live-assist.service.ts suggests statements and follow-up questions. For external information, intent-detector.service.ts and mcp-agent.service.ts identify needs in the conversation, connection-orchestrator.service.ts manages stdio or HTTP MCP servers, and tool-aggregator.service.ts executes relevant tools. Markdown, links, and structured tool results appear in the MCP Results panel. Once recording stops, summary-generator.service.ts creates an overview, topic-based key points, and concrete actions; the application stores the meeting in its history, exports the full record to Markdown, and can post meeting data to configured workflow webhooks. Before a call, the setup wizard can generate probing questions and a discussion checklist from a meeting description, while Google Calendar integration syncs upcoming meetings.
- A macOS-based salesperson wants live visibility into talk balance, pace, questioning, and long monologues while receiving suggestions for what to say next.
- A project lead needs internal meetings converted into attributed topic notes, concrete action items, and a Markdown record containing the full transcript.
- A researcher, consultant, or customer-success specialist wants information needs detected during a call and answered through configured MCP tools without leaving the meeting view.
- An operations team wants completed meeting data delivered automatically to n8n, Zapier, or a CRM through workflow webhooks.
- A meeting-heavy professional wants AI-generated preparation questions and a dynamic checklist based on the meeting context, alongside upcoming Google Calendar events.
What are this agent's strengths and limitations?
- Combines dual-channel live transcription, contextual assistance, conversation analytics, summaries, and action extraction in one desktop workflow.
- Its MCP agent can detect an information need from the active conversation, invoke tools over stdio or HTTP, and render the result inside the meeting interface.
- Uses local SQLite storage and documents concrete safeguards: a loopback-only API, token enforcement, encrypted credentials and MCP secrets, restrictive file permissions, and log redaction.
- Provides both a complete Markdown export and webhook delivery to n8n, Zapier, or CRMs for downstream automation.
- Full recording is currently macOS-only; Windows and Linux builds cannot record because @videodb/recorder has no binaries for those platforms.
- Transcription and AI features require a VideoDB API key, internet connectivity, and transmission of audio, video, and prompts to VideoDB, so this is not a fully offline system.
- Recordings have a hard two-hour limit. Changing it requires editing MAX_RECORDING_DURATION_MS and rebuilding the application.
- Language behavior depends on VideoDB backend support; an unsupported language_code may fall back to the transcription engine's default.
- The supplied repository metadata says the license is unknown, while the README only shows a badge linking to MIT; adopters should verify the actual license file and terms.
How do you install or deploy this agent?
Full recording requires macOS 12 or later. Run:
curl -fsSL https://artifacts.videodb.io/call.md/install | bashLaunch Call.md from Applications or Spotlight, grant Microphone and Screen Recording permissions, and register with a VideoDB API key obtained from the VideoDB Console. For development, install Node.js 18+ and npm 10+, then run:
git clone https://github.com/video-db/call.md.git
cd call-md
npm install
npm run rebuild
npm run devHow do you use this agent?
Open the application, enter a VideoDB API key, and confirm that Microphone and Screen Recording permissions are enabled. Click “New Meeting” to begin; Call.md displays separated live transcripts, assistance, and conversation metrics, then generates the overview, topic notes, and actions when recording ends. To add external tools, open Settings → MCP Servers, select Add Server, choose stdio or http, configure the connection, and click Connect. Set the language for the next recording under Settings → Transcription. After a meeting, review it in recording history, export it as Markdown, or configure a workflow webhook for downstream delivery.