GhostMinutes

GhostMinutes MCP — Audio Transcription for Cursor, Claude, ChatGPT & every MCP client

Accurate meeting transcription through the Model Context Protocol (MCP): speaker labels, multilingual coverage, and hosted Whisper-class models without running GPUs locally.

npm version License: MIT MCP compatible GitHub stars

GhostMinutes MCP is an audio transcription MCP that pairs beautifully with assistants: paste a URL, poll status, summarize transcripts, and inspect credits without leaving your IDE.

Cursor

Use this Cursor MCP server snippet inside Cursor → Settings → MCP → JSON editor:

{
  "mcpServers": {
    "ghostminutes": {
      "command": "npx",
      "args": ["-y", "ghostminutes-mcp"],
      "env": {
        "GHOSTMINUTES_API_KEY": "gm_replace_with_your_key"
      }
    }
  }
}

Removing GHOSTMINUTES_API_KEY keeps you on the anonymous tier while you evaluate meeting transcription workflows.

Claude Desktop

Drop the following Claude MCP server stanza into claude_desktop_config.json:

{
  "mcpServers": {
    "ghostminutes": {
      "command": "npx",
      "args": ["-y", "ghostminutes-mcp"],
      "env": {
        "GHOSTMINUTES_API_KEY": "gm_replace_with_your_key"
      }
    }
  }
}

Restart Claude Desktop after edits so stdio transports reconnect cleanly.

Continue

Continue reads the same stdio definition:

{
  "experimental": {
    "modelContextProtocolServers": [
      {
        "transport": {
          "type": "stdio",
          "command": "npx",
          "args": ["-y", "ghostminutes-mcp"],
          "env": {
            "GHOSTMINUTES_API_KEY": "gm_replace_with_your_key"
          }
        }
      }
    ]
  }
}

Swap in absolute paths only if npx is unavailable inside your sandbox.

ChatGPT custom connector

ChatGPT custom connectors vary by rollout region. Where MCP-over-stdio is unavailable, proxy through your MCP host or paste transcripts retrieved via get_transcript. When stdio bridging is supported, mirror the Claude snippet:

{
  "command": "npx",
  "args": ["-y", "ghostminutes-mcp"],
  "env": {
    "GHOSTMINUTES_API_KEY": "gm_replace_with_your_key"
  }
}

Document any gateway ports carefully—secrets stay local while transcripts upload through GhostMinutes APIs.

Generic HTTP

This npm package exposes stdio MCP using @modelcontextprotocol/sdk. Generic HTTP integrations should terminate TLS upstream (Streamable HTTP lives on GhostMinutes infrastructure separately). Point automation at:

GHOSTMINUTES_API_URL=https://ghostminutes-backend-production.up.railway.app/api

Then attach whatever MCP gateway your ChatOps stack expects.

---

Why GhostMinutes MCP

• Speaker diarization differentiates voices for meeting transcription summaries.
• 99+ languages cover global teams without swapping Whisper MCP forks.
• Hosted infrastructure avoids shipping massive models with every developer laptop.
• Free starter credits keep hobbyists unblocked while Pro unlocks heavier audio loads.
• Model Context Protocol native tooling means one install for Cursor, Claude, Continue, or any spec-compliant host.

Compared with a barebones Whisper MCP, GhostMinutes MCP layers account management, credits, summarization helpers, and consistent job tracking.

Advanced workflow notes

Pair transcribe_audio_url with get_transcription_status whenever assistants cannot await five-minute synchronous polls—most Cursor MCP server automations prefer this split because it mirrors asynchronous CI patterns.

Need deterministic summaries? Fetch transcripts first (get_transcript), stash them in version control redacted snippets, then feed curated excerpts into summarize so Model Context Protocol audit trails stay reproducible.

Meeting transcription pipelines benefit from tagging URLs with HTTPS endpoints only; GhostMinutes rejects flaky redirects early to save Whisper MCP troubleshooting loops.

Because MCP-compatible hosts sandbox stdio pipes differently, keep logs on stderr inside downstream forks—GhostMinutes MCP itself never prints informational noise on stdout.

When benchmarking GhostMinutes MCP versus generic Whisper MCP integrations, capture wall-clock latency for both anonymous and authenticated tiers separately—credit-aware queues behave differently under burst uploads.

Power users orchestrating ChatGPT custom connector gateways should terminate TLS at their proxy while forwarding Authorization: Bearer gm_* headers verbatim to preserve quota attribution.

Finally, bookmark examples/ for ready-to-paste snippets tailored to Cursor, Claude Desktop, and Continue—each file mirrors best practices validated against @modelcontextprotocol/sdk v1.x transports.

Available tools

Tool	Description	Parameters	Auth required
transcribe_audio_url	Submit a public audio/video URL for asynchronous transcription	url, optional language	No (optional key)
transcribe_audio_url_sync	Submit a URL and poll up to five minutes	url, optional language	No (optional key)
get_transcription_status	Poll transcription job metadata	job_id	No (optional key)
list_transcripts	List saved jobs for the signed-in workspace	(none)	Yes
get_transcript	Download a transcript payload by id	id	Yes
delete_transcript	Remove a transcript	id	Yes
summarize	Hosted AI summarization via /api/ai/chat	text, optional focus	Yes
get_credits	Inspect remaining GhostMinutes quota	(none)	Yes

Anonymous traffic is perfect for prototypes; authenticated tiers unlock persistence and collaboration.

Pricing & credits

Tier	Monthly credits	Typical best for	Notes
Anonymous	10 credits / month / IP	Quick experiments without signup	Ideal for Whisper MCP comparisons
Free (signed in)	10 credits / month	Individuals testing meeting transcription	Includes dashboard access
Pro	300 credits / month ($9.99)	Agencies & power users	Predictable throughput

Rule of thumb: 1 GhostMinutes credit ≈ 6 minutes of audio. Actual usage varies with diarization complexity and retries.

Need more? Visit ghostminutes.com for the latest bundles and promotions.

Use cases / example prompts

1. Exec recap — “Call transcribe_audio_url_sync on this all-hands MP4, then summarize with focus CEO decisions only.”
2. Support QA — “List transcripts via list_transcripts, fetch #472 with get_transcript, highlight escalations.”
3. Localization smoke test — “Send transcribe_audio_url with language es, poll using get_transcription_status, verify diarized segments.”
4. Budget guardrail — “Run get_credits, refuse further audio transcription MCP jobs if fewer than two credits remain.”

These prompts showcase how GhostMinutes MCP complements Cursor MCP server workflows without bouncing between terminals.

FAQ

Do I need to sign up?

Anonymous usage works for lightweight Whisper MCP comparisons. Sign up when you want authenticated tools (list_transcripts, summarize, higher reliability dashboards).

Is my audio kept?

GhostMinutes stores transcripts according to the production privacy policy linked from ghostminutes.com. Prefer ephemeral URLs or rotate keys after audits.

What languages are supported?

The backend mirrors Whisper-class multilingual coverage—omit language for autodetection or pass ISO 639-1 codes for forced alignment.

How do I get more credits?

Upgrade inside the GhostMinutes billing portal or rotate teams onto Pro seats when meeting transcription volumes spike.

Can I self-host?

This repo focuses on the hosted GhostMinutes stack. Enterprise self-hosting questions should route through Rocketech support channels.

Comparison: GhostMinutes MCP vs generic Whisper MCPs

Dimension	GhostMinutes MCP	Typical Whisper MCP
Speaker diarization	First-class job metadata	Rarely bundled
Credits & auth	Built-in GhostMinutes API	Bring-your-own infra
Summaries	Native summarize tool	Usually manual prompts
Meeting transcription UX	Designed around job lifecycle	Often single-shot
Claude / Cursor parity	Same config across hosts	varies per template

Configuration / env vars

Variable	Required	Description
GHOSTMINUTES_API_KEY	No	API key (gm_<32-char-hex>). Without it you get anonymous limits.
GHOSTMINUTES_API_URL	No	Override API base if you proxy through a private edge. Defaults to Railway production

Reference server.json for the canonical MCP Registry manifest submitted to registry.modelcontextprotocol.io.

Roadmap

• Batch URL ingestion for playlist-style meeting transcription imports.
• Richer structured outputs for diarization segments and per-speaker timelines.
• Tooling for redacted exports (PII masking) prior to summarization.
• Expanded automated tests covering edge cases from the hosted API.
• Additional documentation for IDE-specific authentication patterns.
• Companion observability dashboards describing MCP traffic separately from HTTP dashboards.

Contributing

Issues and PRs are welcome in Rocketech-Software-Development/ghostminutes-mcp. Please include repro steps, sample URLs (sanitized), and expected MCP tool outputs. CI runs npm run build and npm test on every push to main.

License

MIT — see LICENSE.

GhostMinutes MCP remains independent from upstream Whisper releases; Rocketech ships compatibility-tested bundles so Cursor MCP server upgrades stay boringly predictable.

Commercial support, custom SLAs, and private Model Context Protocol gateways are available—email the GhostMinutes team via the website contact form.