MCP server for AI-driven music score generation and manipulation
mcp-score has 18 well-organized tools with consistent naming (verb_noun pattern: connect_to_*, read_*, get_*, add_*, set_*, etc.). Tool descriptions are present and moderately detailed (average ~180 chars, within the 10-1024 baseline). Input schemas are visible and typed. However, several critical gaps limit the score: (1) output schemas are NOT documented, the code shows tools returning dict[str, Any] with no structured schema definition for LLM consumers; (2) no error handling guidance, tools lack recovery instructions or categorization (retryable vs fatal); (3) descriptions lack dependency hints and context for when to choose one tool over another; (4) parameters have types but descriptions are sparse (e.g., 'WebSocket host (default: localhost)' is minimal); (5) tool annotations (readOnlyHint, destructiveHint, idempotentHint) are absent despite having mixed risk profiles (READ_ONLY, WRITE, REVERSIBLE). These gaps are typical of alpha-stage servers but prevent confident production use.
Add a chord symbol in the live score.
Add a rehearsal mark in the live score.
Connect to a running Dorico instance via its Remote Control API. Dorico 4+ has a built-in WebSocket server (no plugin needed). The port is configurable in Dorico's preferences.
Connect to a running MuseScore instance. The MuseScore MCP Score Bridge plugin must be running.
Connect to a running Sibelius instance via Sibelius Connect. Sibelius 2024.3+ has a built-in WebSocket server (no plugin needed). Requires Sibelius Ultimate tier. The port is configurable.
Disconnect from Dorico.
Output schemas not documented. All tools return dict[str, Any] with no structured response schema visible to LLMs. Agents cannot plan downstream operations or extract chaining IDs (e.g., score ID, measure number) without guessing field names.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk metadata is tracked in the rubric but not exposed to clients. Agents cannot determine which tools are safe to retry or which have irreversible side effects without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 64 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Disconnect from MuseScore.
Disconnect from Sibelius.
Get information about the currently open score. Requires an active connection — use connect_to_musescore, connect_to_dorico, or connect_to_sibelius first.
Read the content of a specific measure and staff from the connected score. Works best with MuseScore; Dorico and Sibelius return limited data.
Get properties of the current selection in the connected score application. Returns information about whatever is currently selected: - **MuseScore**: Returns cursor position info (measure, beat, staff). - **Dorico/Sibelius**: Returns properties from the Remote Control API's ``getproperties`` message — names, types, and values of all properties on the selected items. This is the closest the WebSocket API gets to "reading" score data. Requires an active connection.
Check if the connected score application is responsive. Works with any connected application (MuseScore, Dorico, or Sibelius). Does NOT auto-connect — returns an error if not already connected.
Read musical content from a range of measures in the live score. Returns notes, rests, and musical elements in the specified range. Works best with MuseScore; Dorico and Sibelius return limited data through the Remote Control WebSocket API.
Set a barline type in the live score.
Set the key signature in the live score.
Set the tempo in the live score.
Transpose a passage by a number of semitones in the live score.
Undo the last action in the connected score application.
No error handling guidance or categorization. Tools do not document what errors can occur, whether they are retryable, or what recovery steps the agent should take. For example, 'connect_to_musescore' fails silently if the bridge plugin is not running.
Parameter descriptions are minimal. Most parameters have 1-2 sentence descriptions without context on format, range, or constraints. E.g., 'Measure number (1-indexed)' lacks guidance on valid range or what happens if out of bounds.
No dependency documentation. Tools like 'add_live_rehearsal_mark' depend on an active connection, but descriptions do not explicitly state prerequisites or guide multi-step setup (e.g., 'call connect_to_musescore first').
Limited parameter constraints. Enum values (e.g., 'barline_type' accepting only 'double', 'final', 'startRepeat', 'endRepeat') are documented in descriptions but not as formal JSON Schema enums. This forces LLMs to parse text constraints rather than leveraging structured validation.