Everyone in a friend group gets their own AI, called a bee. Behind all the bees sits one shared brain, the hive. The hive quietly learns what everyone is up to, remembers it, and connects people when it makes sense. It also knows how to keep a secret: it decides what is okay to pass between friends and what should stay put.
A hosted instance runs here:
- Bee chat: https://hive-demo.onrender.com/chat
- Hive dashboard: https://hive-demo.onrender.com/
It boots empty — nothing is pre-seeded. The operator adds members on the dashboard, and people pair with an invite code before there's anything to see. It runs on a free host that falls asleep when idle, so the very first load can take about 50 seconds to wake up.
flowchart LR
A[Member] --> H[Hive brain]
B[Member] --> H
C[Member] --> H
H --> D[Dashboard]
H -.->|nudges| A
H -.->|nudges| B
H -.->|nudges| C
Each person only ever talks to their own bee. The hive sits in the middle, builds a memory of the whole group from real conversations, and reaches out when it spots something worth a nudge.
Hive is two sites. One is for the people in the group, one is for whoever runs it.
| Bee chat | Hive dashboard | |
|---|---|---|
| Who it is for | a member | the operator |
| What you do | chat with your bee | watch the hive think |
| Where | /chat |
/ |
Where a person talks to their bee.
- Paste the invite code (
BEE-XXXX) the operator gave you to pair. - Type a message and the bee replies live. Everything you say flows into the hive.
- Ask about someone else and watch what it will and will not tell you.
The top-left Profile picker is a multi-account switcher: paste another member's code to add their profile and switch between people from one browser.
The operator's window into the brain. Tabs down the left side:
- Knowledge graph: everything the hive has learned, drawn as a live map of people and things.
- Proactive: who the hive wants to reach out to and why. Hit Find connections to make it hunt for new ones.
- Disclosures: a receipt for every time info crossed between people, including what it held back.
- Polls: ask the whole group something anonymously and get one answer back.
- Members: the people, their invite codes, and whether each bee is online.
- Channels and Settings: connect Telegram / Discord, and pick which AI model does which job.
Every message a bee receives runs down the same little assembly line:
flowchart LR
T[chat turn] --> E[extract facts with an LLM]
E --> G[(knowledge graph in SQLite)]
G --> R[recall for replies]
G --> O[find connections]
O --> DA{disclosure agent}
DA -->|safe to share| N[nudge a person]
DA -->|private| X[held back]
Nothing is hardcoded and nothing is pre-seeded. The hive starts empty; the graph, the disclosures, and the nudges all grow out of the real conversations members have with their bees.
Say Devon quietly tells his bee he is job-hunting and asks to keep it private — especially from anyone connected to his work. That fact now lives in the hive. Then Sam, who is hiring a backend engineer, asks his bee, "know anyone good who might be open to a new role?"
Devon's bee -> hive learns "job-hunting, confidential" (private to Devon)
Sam asks -> hive runs the disclosure agent -> withhold
Sam hears -> nothing about Devon (secret kept)
The clever part: the hive still knows Devon is a perfect match, and it will happily share his climbing hobby to connect the two as friends — it just refuses to surface the job hunt to someone in his work orbit. Same person, one fact shared, one withheld. The disclosure agent runs every time knowledge would cross between people; it can share, partially share, or withhold, fails safe (anything goes wrong → withhold), and writes down its reasoning every single time, which is what fills the Disclosures tab. (The screenshot below is this exact case, live.)
- Memory with a timeline. Facts are stored with a validity range. When something changes, the old version is marked outdated instead of deleted, so there is history.
- Contextual-integrity disclosure. The share-or-withhold judgment above, fully audited.
- Proactive reach-outs. A heartbeat looks for introductions worth making and things you would want to know, with cooldowns and quiet hours so it does not get annoying.
- Ask your network. Post an anonymous question to the group and get one synthesized answer back — with an anonymity floor so a lone reply is never attributable.
- Real-world errands + web lookup. The bee can search the web (no API key needed — keyless by default, Exa if you add a key) and go find things you mention wanting.
- Reminders. Ask to be reminded of something later and the bee will message you when it's due.
- Reach people anywhere. Web chat always works. Telegram and Discord plug in too.
- Any model. Anthropic, MiniMax, OpenAI-style endpoints, or local Ollama. Three jobs (chat, extraction, social) can each run on a different one.
- Keys stay safe. Provider keys are encrypted on disk, and bees never hold them.
Ask the whole group a question anonymously and the hive gathers replies and synthesises one answer:
pnpm install
pnpm dev # dashboard on :5173, bee chat on :5174The hive comes up empty. Add a model key in Settings, add a member in Members, then open the chat and paste that member's code to pair. From there, real conversations populate the graph, disclosures, nudges, and polls. The full walkthrough is in docs/SETUP.md.
- docs/ARCHITECTURE.md: how the graph, disclosure, and proactive systems work.
- docs/HOSTING.md: the single-container build and the current Render deployment.
- docs/SETUP.md: operator and member setup, including the channel bots.