Try it yourself
Each door opens the real app in a specific state. Nothing is a recording; with the server's Qloo key every call is live.
Receipts from a live run
Run one session now and see every Qloo call it made, whether it was answered live, and the hash chain that seals the trace. Then verify the chain on the server.
- The Qloo calls of the session appear here, in the order the agents made them.
Reach inside your assistant
Reach is also an MCP server, so Claude, ChatGPT or any agent can ask it for a session in the middle of a conversation: "I can't sleep, I love Hayedeh." The assistant gets the plan, the evidence and a link that plays it.
/mcp- Claude: Settings, Connectors, Add custom connector, paste the address.
- ChatGPT: Settings, Connectors, Create (developer mode), paste the address.
compose_sessionfavourites and a need: the full council, Qloo picks with why, the evidence grade, a link to playafter_the_musica place, a table and a book near them, never a barexplain_evidencethe graded research behind a need, with sourcesreach_vs_chatbotthe measured comparison belowAgainst a plain chatbot
The same ten people (Persian, Mandarin, Afrobeats, K-pop, Turkish, Bollywood, Mexican, techno, a mix, and lively favourites who need sleep) asked three systems for music. Picks of the person's own favourites are not counted.
What leaves the browser
The Qloo key stays on the server. Mood from face and voice is measured in the browser, only as numbers, and only with consent. The full privacy page says where every piece goes.
Sent to Qloo
- Entity ids of the favourites the person typed
- Genre and place-category tags chosen by the session's protocol
- An age bucket (child or adult), for content ratings
- A city, for the cultural lens and for places near them
Never sent
- Their words: the conversation stays with Reach and the conversation model the server runs (a local one by default)
- Face and voice: no camera image or voice recording reaches Qloo or is stored by Reach (talking to the chat uses the browser's own speech recognizer)
- Mood, health or the goal itself: the goal only chooses which tags are asked for
- Names: the person's name is never part of a Qloo request
Where each criterion shows
Technological implementation
Eight Qloo surfaces in one agent: entity search, artist insights with protocol tag filters, the location signal for the cultural lens, tag insights for texture, demographics, podcasts, films and shows, places with category filters, and books. Explainability drives every "close to" label. Parameters are allow-listed before a request is built, calls are rate-spaced and cached, and the trace is hash-chained. It is also an MCP server, so any assistant can call it.
See the live callsDesign
A finished product, not a form: a 3D start page where every capability is a world you can fly to, a frequency instrument you can play, a 3D film introduction, a 3D night map of the person's city, push-to-talk with ducking, hand gestures, fourteen languages including right-to-left Persian and Arabic, an installable app on a phone, and a mode for people who can't hear.
Open the appPotential impact
Music lowers anxiety and pain (81 trials, 5,576 people with cancer, Cochrane 2021), but generic calm playlists get switched off. Reach makes the evidence-based session out of music the person loves, in their own culture, with safety rules for children, autism and hearing loss.
The care team's viewQuality of the idea
Taste as the carrier of a clinical protocol: the protocol decides the tempo path, pulse band and safety envelope; Qloo decides what music carries it. A council of agents argues, three decision rules vote, and a gate refuses when the evidence or Qloo is missing.
The comparison