Skip to content

ROADMAP.md

Latest commit

 

History

History
41 lines (31 loc) · 2.31 KB

ROADMAP.md

File metadata and controls

41 lines (31 loc) · 2.31 KB
Raw

Roadmap

MusicContext wants to be the music-direction layer AI agents reach for. That means depth in reasoning and honesty about uncertainty, not a longer feature list.

Now (0.1.x)

  • Harden what exists: more real-world media shapes (vertical video, long-form, variable frame rate, multi-track audio).
  • Verify Codex end to end and document the result (docs/agents/codex.md).
  • Tighten the heuristics with a labelled evaluation set instead of intuition.

Next (0.2)

  • Better speech handling. Optional local transcription (opt-in, local model, no upload) so voiceover timing is detected rather than estimated without asking the user for a transcript.
  • Per-section music. Multiple cues per video with planned transitions between them, rather than one track.
  • Optional vision assist. A pluggable visual analyzer for shot type, text/logo appearance and product moments, behind the same provider-style interface, with graceful degradation when absent. Reveals become detected.
  • Stem-aware mixing. When a provider supplies stems, duck and thin arrangements per stem instead of per track.
  • Agent evaluation suite. A scored benchmark for tool choice, timestamp accuracy, licensing correctness and injection resistance, run in CI.

Later

  • Real provider integrations, contributed behind the existing interfaces, each with its own contract tests and a documented data-handling statement.
  • Beat-accurate conform. Nudging cuts to the music (an edit-side change) rather than only music to cuts, exposed as an explicit opt-in because it alters the user's edit.
  • REST service mode with the same sandbox and bounded output as the MCP server, for queue-based deployments.
  • DAW/NLE handoff. Export markers and cues as a format an editor can open (e.g. marker lists, AAF/OTIO investigation).
  • Loudness targets per platform, verified against each platform's published normalization behaviour.

Explicit non-goals

  • Becoming a video editor, a DAW or a streaming service.
  • Shipping a bundled music catalogue or a generation model in this package.
  • Telemetry, or any default that sends media off the machine.
  • Guessing. If MusicContext cannot measure something, it will keep saying so.

Have a use case that does not fit? Open an issue describing the video and the musical decision you need made.