Skip to content

AI Assistant ​

Subcast can expose a controlled local entry point for AI assistants. Once enabled, Codex, Claude Desktop, and other MCP clients can use Subcast locally: import media, read processing status, start transcription or AI Insights, and export timestamped evidence packs.

This entry point is off by default. Once you enable it, Subcast remembers the choice on this device and refreshes the local credential each time the app starts.

1. Enable local access ​

Open Subcast, go to Settings → AI Assistant, then click Enable local AI Assistant access.

Subcast writes a temporary access profile inside the local app-data directory. It is readable only by the current user and is refreshed on each launch. Do not paste desktop session tokens or access-profile contents into chat; supported agents should use MCP or the Subcast skill to read the authorized local profile.

2. Install the Codex skill ​

From the source repo, run:

bash
pnpm skill:install

It installs the Subcast skill to ~/.codex/skills/subcast. To intentionally replace an older install, run:

bash
pnpm skill:install -- --force

Then ask Codex for tasks like:

text
Use Subcast to transcribe /path/to/video.mp4
Use Subcast to export a media pack for this video

3. Configure MCP ​

For the packaged app, prefer the bundled MCP launcher:

json
{
  "mcpServers": {
    "subcast": {
      "command": "/Applications/Subcast.app/Contents/Resources/subcast-mcp"
    }
  }
}

For source development, build the desktop entry first:

bash
pnpm build:desktop:main

Then point your client at the source entry:

json
{
  "mcpServers": {
    "subcast": {
      "command": "node",
      "args": ["/absolute/path/to/subcast/desktop-dist/subcastMcp.js"]
    }
  }
}

4. What's in a media pack ​

The basic archive pack exports:

  • manifest.json: redacted run metadata;
  • transcript.md: timestamped transcript;
  • subtitles.srt: subtitles from the same cue source;
  • chapters.md: cached AI Insights chapters when available, otherwise a missing-state note;
  • summary.md: cached AI summary when available, otherwise a missing-state note;
  • sources.json: cue-level evidence map;
  • deliverable.md: a short handoff note for downstream agents.

For clip ideas or meeting notes, run matching-language AI Insights in Subcast first, then export the creator-brief or meeting-notes recipe.

Safety boundary ​

  • Import only absolute local paths explicitly provided by the user.
  • Authorization remains enabled on this device until you revoke it in Settings → AI Assistant.
  • Subcast still uses a fresh local credential on each launch; do not paste credentials into chat.
  • Agent responses expose hash prefixes, readiness, and next actions by default.
  • Do not expose original paths, full hashes, desktop tokens, prompts, or raw model output in chat.

Released under the Apache-2.0 License