Runs offline · your music stays on your machine
Turn audio, MIDI and production notes into a searchable project library. Measure tempo, key and loudness, track unfinished work, and ask questions with answers linked tothe original files.
Architecture & evaluation
Actual workspace · Generated demo library
The problem
Every project, every bounce, every version in one organized place — with search across the whole catalog and files that move on disk when you rename a project.
Track metadata, open tasks and a readiness checklist for each song make unfinished work visible across the library.
Mark a track as your reference and every mix is measured against it — signed differences in loudness, tone and length, so "it sounds quieter" becomes a number.
Session notes, a decision log, and preferences that persist. Record what you chose and it comes back later, quoted, with the file it came from.
Your library, measured
Tempo, key with major or minor, peak and RMS level, integrated loudness to the broadcast standard, spectral brightness, beat positions, and for MIDI the notes, pitch range and instruments. All of it computed locally, on your machine.
| File | Project | BPM | Key | Loudness | Length | Status |
|---|---|---|---|---|---|---|
| midnight_drive.wav | Neon Horizon / Midnight Drive | 117 | E | −18.3 | 8.0s | In Progress |
| afterglow.wav | Neon Horizon / Afterglow | 117 | E | −18.4 | 8.0s | Idea |
| afterglow_master.wav | Neon Horizon / Afterglow | 117 | E | −18.4 | 8.0s | Ready |
| lofi_sketch.wav | Lo-Fi Study | 76 | C | −18.9 | 7.0s | Idea |
| cue_draft.wav | Client Cue 03 | 99 | A | −19.9 | 6.0s | Needs Work |
Sample from the demo library SessionIQ generates on first run.
Mix vs reference
Mark any track as your reference and every mix in the project is measured against it — signed differences in loudness, tone and length, available alongside the source files.
afterglow.wav vs afterglow_reference.wav
Ask it anything
Ask across one song or the entire catalog. Follow-ups work — “what about its key?” — and every answer arrives with the files it came from, so you can check it.
Counts, comparisons and superlatives are computed from the real metadata rather than guessed at, which is why "which track is the loudest" gets an exact answer instead of a list of likely candidates.
How it is measuredWhich track is the loudest?
Client Cue 03/reference.wav — integrated loudness −15.5 LUFS, the highest in the library. The quietest is afterglow_reference.wav at −25.0.
Which tracks are the fastest?
Four tracks tie at 117.5 BPM: midnight_drive.wav, afterglow.wav, afterglow_master.wav and afterglow_reference.wav. The slowest is lofi_sketch.wav at 76 BPM.
Answers shown are the computed result over the generated demo library. Every number is checked against the cited files before an answer is shown; one that fails is withheld.
Screens
Insights
Projects ranked by how close they are to finished, with the stalled ones surfaced instead of buried. Search finds files by what they sound like rather than what they are called, and the similarity engine shows you why two tracks were paired — tempo, brightness, loudness, length and key, broken out separately.
Studio
A fingerprint built from your own catalog rather than typed in — the tempo you actually work at, the keys you keep returning to, how loud your mixes land. Your preferences come back on the next session, and the timeline is a history of what entered the studio and when.
Under the hood
The system separates ingestion, analysis, retrieval, answering and validation. The repository includes an evaluation harness with labeled queries, recorded results and documented limitations. The measurements below can be reproduced from that harness.
Audio, MIDI, notes, artwork
Tempo, key, loudness, brightness
Lexical, metadata, optional vectors
Grounded, tool calls for exact counts
Citations and numbers checked
The running system, not a diagram — each stage reports whether it is active.
Your own agent can ask the same questions. SessionIQ speaks the Model Context Protocol, so Claude Desktop, Cursor or an agent you write can query your library through the identical tool layer the numbers below are measuring.
MCP setupStated plainly: these come from one synthetic corpus of seventeen generated files and thirty hand-written questions. That is enough to compare retrieval configurations and far too small to be a benchmark. The evaluation does not score the wording of generated answers, and it does not measure hallucination. Both limits, and every case retrieval gets wrong, are written down in the methodology.
Methodology and limitsGetting it
Core analysis and library management run on your machine, with no account required. Use the offline assistant or configure a local or cloud model for generated responses.
One command starts it, with a demo library already built in so you can look around before pointing it at your own files. Semantic search and voice transcription are optional local add-ons, and the assistant works with no model configured at all.
$ docker run --rm -p 8080:8000 sessioniq
Your library http://localhost:8080
Demo songs 17 files, 5 projects
Assistant offline / rules engineWhat you need
Docker, if you want the one-command route. Prefer running it from source? Python 3.11+ and Node 22.18+ — the README walks through it, including optional local Ollama, ChromaDB and Whisper.