Three hymns, sung on the beat. Assembled from sixteen takes, measured, then heard.

A score clock

Every syllable gets its pitch, its onset and its length from the hymn's own arrangement, on the same timeline the piano plays. Held notes, rests and breaths included.

A sound check first

Before a song is rendered, the same voice sings a sixteen-word calibration phrase at the song's tempo. Its timing footprint, per consonant group and for held against short notes, says how many takes the song needs and which words are risky.

Sixteen takes on a rented GPU

A score-conditioned singer renders each hymn sixteen times on a rented A40 through offrig, so the local card stays free. Each phrase is sung with a second of silence in front, then trimmed, which keeps the model's start-up noise out.

One take per phrase

A local listener transcribes every phrase of every take without being given the lyrics, and pitch is measured note by note. Each phrase comes whole from the take that was understood best and sung most in tune.

Placed by warping, not cutting

Each phrase is time-warped onto the score so every vowel lands on its beat, with nothing replayed or skipped inside the phrase. Takes meet only at the rests between phrases.

Two gates, two instruments

Timing: an energy detector dates each vowel, a singing forced aligner checks it, and a syllable fails only when both say it is off. Pitch: every note is tracked against the score in cents.

A person listens last

The listening review: press M wherever something sounds wrong, and a report joins each mark to the take, the join and the gates at that moment. A listener's ear settles where something sounds off; naming the cause weighs more from training. Its first marks found two defects no gate could see, both fixed at the root. It is moving into the cockpit next.

Reading

Real MIDI sheet music with deep musical annotations — parsed, analyzed, and explained. Not hand-written approximations.

Hearing

Six audio engines play through your speakers so the humans in the room become the AI's ears.

Seeing

Piano roll renders as SVG the model can read back. Interactive guitar tablature. Browser cockpit with visual keyboard.

Remembering

Practice journal persists across sessions. Next time, the AI reads its journal and picks up where it left off.

Singing

Whole hymns sung by a score-conditioned singer and placed on the piano's clock (listen above), plus vocal tract synthesis with 20 voice presets and a sing-along mode with solfege, contour and syllable narration.

56 MCP Tools + 4 Prompts

Learn, play, sing, build, score — browse songs, transpose keys, mute hands, preview teaching cues, render piano rolls, write annotations, and journal reflections.

A Training Dataset to Match

jam-actions-v0 — a public corpus of 57 multi-turn MCP tool-use traces over classical piano, with a 7-axis release gate and cold-start reproducibility. CC-BY-SA-3.0-DE.

57 records · 4 piano pieces

Public subset 0.6.0: four piano-midi.de arrangements (Bach, Beethoven, Debussy, Mozart), 45 train and 12 held out. Version 0.6.0 withdrew 58 records from four songs whose files had no established arrangement licence. The 16-record canonical baseline was measured on the earlier composition.

7-axis release gate

Absolute floor, margin compound, tool-use rate, correct-after-tool, misinterpretation count, stratum floor (all blocking); enriched-vs-non reporting (informational). Admits a ceiling-saturated bucket so trivial wins do not dilute harder strata.

Reproducible in under a minute

pnpm install, run the checksum verifier, run the release-gate CLI against the canonical baseline. .gitattributes pins LF for sha256 + dataset tree so it works on Windows native, macOS, Linux, and WSL.

Slice 22 baseline PASSES

The revised gate (axes 2 + 6 with ceiling_saturated_pass) admits the canonical baseline cleanly. The Slice 19 baseline still FAILS — kept as a regression diagnostic so the gate has teeth.

Cited and licensed end-to-end

CITATION.cff, Zenodo deposition metadata (13 fields, ISO 639-3 language code), CC-BY-SA-3.0-DE preserved across the MIDI arrangements (Krueger) and the annotations + traces + evals (mcp-tool-shop-org).

Honest about provenance

Version 0.6.0 withdraws Chopin's Nocturne Op. 9 No. 2 and Prelude Op. 28 No. 4, Beethoven's Pathétique II, and Schumann's Träumerei. Those files came from midiworld.com and bitmidi.com and carry no established arrangement licence. The remaining 57 records are byte-identical to 0.5.0.

Fine-tuning receipts, three arcs

Preregistered fine-tunes on sealed baselines: v0 was an honest negative (0.661 → 0.601); v1 hit +0.202 but missed its frozen 13/16 bar by one; B-1 cleared its ex-ante bar (0.678 → 0.890, 29/36 wins). The five seed adapters are withdrawn: they were trained on records that include the four songs 0.6.0 removed. The measurements stand as recorded.

Verify the package

git clone https://github.com/mcp-tool-shop-org/ai-jam-sessions.git
cd ai-jam-sessions && pnpm install

# 274 checksum entries; ~2 seconds.
pnpm exec tsx scripts/verify-public-package-checksums.ts

Reproduce the canonical PASS

pnpm exec tsx scripts/check-release-gate.ts \
  datasets/jam-actions-v0-public/evals/\
slice21-fair-e3-baseline-results.json
# → "Aggregate: PASS" (exit 0)

Read the dataset card

# Full HF-format dataset card + YAML frontmatter
open datasets/jam-actions-v0-public/README.md

# Zenodo deposition metadata (13 fields)
cat datasets/jam-actions-v0-public/zenodo-metadata.json

Cite it

# Citation File Format (CFF)
cat datasets/jam-actions-v0-public/CITATION.cff

# Bernd Krueger (piano-midi.de) for the MIDI arrangements.
# mcp-tool-shop-org for the annotations, traces, and evals.
# Both under CC-BY-SA-3.0-DE.

Install globally

npm install -g @mcptoolshop/ai-jam-sessions

Claude Desktop config

{
  "mcpServers": {
    "ai_jam_sessions": {
      "command": "npx",
      "args": [
        "-y", "-p",
        "@mcptoolshop/ai-jam-sessions",
        "ai-jam-sessions-mcp"
      ]
    }
  }
}

Play a song

# Play Fur Elise at 70% speed
ai-jam-sessions play fur-elise --speed 0.7

# View piano roll as SVG
ai-jam-sessions view autumn-leaves --measures 1-16

# Sing along with piano
ai-jam-sessions sing imagine --with-piano

Guitar tablature

# View interactive guitar tab
ai-jam-sessions view-guitar greensleeves

# List tunings and voices
ai-jam-sessions list --genre folk