Kora vs AI Music Tools
Direct answer: Kora does not generate music. It has no generation model, no text-to-audio endpoint, and no connection to any music generation service. Kai, the assistant built into Kora, works on the metadata of your work — project names, due dates, delivery statuses, contact follow-ups — and on text you explicitly hand it. It never reads, analyses or uploads your audio.
If you arrived looking for a tool that writes music, Kora is not it, and no amount of positioning will make it it. If you arrived worried that installing Kora means an AI ingesting your catalogue, the rest of this page is the specific answer.
Is Kora an AI music generator?
No. Kai’s capability registry is a closed list of 44 enumerated actions, all organisational — create a task, schedule a follow-up, summarise a brief, rank priorities. There is no action that produces audio, because no such action exists in the list.
Two structural facts back that up rather than asserting it:
- The only audio recording path in the entire assistant is the microphone for dictation. Every outbound endpoint is a text or transcription API.
- The server injects “You generate text only” into every prompt, and Kai’s hard-prohibition list contains exactly two entries: delete anything and move money.
What does Kai actually read?
Precisely enumerated, because “it sees your workspace” is not an answer:
- Page and UI context — where you are in the app
- Names only for projects, albums, contacts, releases and opportunities
- Album overviews: name, due date, track count
- Top projects: name, due date, priority, status, score
- Tasks due today: title, priority, dates, and the names of linked project, album or contact
- Contact follow-ups: name, days until due, roles
- Your current mood setting and release readiness scores
- Your composer profile, learned memory facts, and active beliefs with confidence values
- The last 8 chat turns, each truncated to 220 characters
- A preview of whatever entity is open — project notes cut at 180 characters, album brief at 200
The boundary is enforced by construction. There is no scrubber stripping sensitive data out of a larger payload; instead a hand-written whitelist decides what is assembled in the first place. Nothing reaches the model that someone did not explicitly add to that list.
What leaves your machine?
Three things are worth naming plainly, because a careful reader would find them anyway:
| Audio | Never. No exceptions, no endpoint, no code path. |
| Export folder path | A real folder path is included with the delivery digest. |
| Note text | Snippets of up to 600 characters from as many as 15 notes travel with the agentic path — and the full text when you explicitly ask Kai to summarise or rewrite something. |
| Recent conversation | The last 8 chat turns, each truncated to 220 characters, accompany every cloud request so Kai can follow the thread. |
The honest one-line version: Kai reads names, dates, statuses and the text you ask it to work on. It never reads, analyses or uploads your audio.
Does Kai work without an internet connection?
Yes — and local is the default, not a degraded fallback. The engine mode defaults to
local, and the Local tier never touches the network at all.
Requests route in tiers: deterministic local handling first, which never leaves the device; a lightweight cloud model for everyday requests; and a stronger model only when you ask for it or when the lighter one flags that it cannot answer well. If a cloud call fails, the chain falls back to the local engine rather than erroring, so you always get an answer.
How do I see why Kai said something?
Every reply carries a “Why · context” control — “Show what Kai knows and why it answered this way.” Opening it shows:
- What Kai is working from — e.g. “Your Composer Profile”, “7 learned memory facts”, “Live studio intelligence (projects, deliveries, releases, relationships)”, “3 active beliefs”
- Why this direction, the tools used, and a priority lens showing the ranking it applied as
Ranked: A → B → C - Evidence — up to four supporting points
- Beliefs Kai used, each with a numeric confidence such as
Confidence 78% · medium
Beliefs are correctable. Each carries “That’s right”, “Not true” and “Don’t use this”, and each reply can be marked “Did this”, “Helpful” or “Off-base”. An assistant that holds a wrong belief about your workflow and cannot be told so is a worse tool than one that holds no beliefs.
Showing confidence as a percentage is a deliberate choice. It invites you to disbelieve Kai when the number is low, which is the correct response.
Frequently asked
Is Kora an AI music generator?
No. Kora has no generation model, no text-to-audio endpoint and no integration with any music generation service. Kai’s capabilities are a closed list of 44 organisational actions, and the server instructs it that it generates text only.
Does Kai listen to or analyse my audio?
No. The only audio path in the assistant is the microphone used for dictation. There is no code path that sends audio files anywhere for analysis.
What data does Kai send to the cloud?
Entity names, dates, statuses, priorities, your composer profile, learned facts, the last eight chat turns, and short note snippets — plus your export folder path with the delivery digest. Full note text is sent only when you explicitly ask Kai to summarise or rewrite it.
Can I use Kai offline?
Yes. The engine defaults to local, and the Local tier never contacts the network. Cloud models handle heavier requests when available, and the chain falls back to the local engine if a cloud call fails.
How do I know why Kai gave a particular answer?
Open the “Why · context” panel on any reply. It lists what Kai was working from, the reasoning direction, the tools used, up to four pieces of evidence, and every belief it applied with a numeric confidence score you can correct.