CrewBrain

Meet Juno and Rex: AI helpers your homeowners and crew can text or talk to. They only see what you allow, nothing goes out without your OK, and they switch off when the job ends.

TL;DR

• A small contractor tells CrewBrain who needs help, by voice, in any language. One card to approve, and a helper goes out by text or QR code.
• Every piece of company knowledge carries a lock. Helpers search as their people in HydraDB, so locked notes never come back. Proven on real query results and by 86 tests.
• Helpers do real work (log a damage photo, request a walkthrough, draft an update for the homeowner), always with the owner in the loop, and they stop the moment the job closes.

Live demo	https://macmini.tail435588.ts.net:10000/crewbrain/demo.html (pick a person, no sign-up; every visitor gets a private copy)
Landing page	https://verkyyi.github.io/agent-creator/
Demo video (3:10)	YouTube · Watch on the landing page · Download 1080p
Devpost	https://devpost.com/software/crewbrain
Built for	Company Brain Hackathon, Menlo Park, Oct 12 2026

Three ideas

1. Say it, get a helper. Sam, the owner, says one sentence in any language, by voice or text: "Brief Bay Roofing's crew for the Oak St job until it closes Friday." The assistant asks only what's missing and shows one compact card: who can use it, what it knows, what it can do, what it can't see, when it ends. Nothing goes live until Sam taps Approve. Links go out by text, email or QR.
2. Only what they should see. Every piece of company knowledge has a lock: Office only, Crew for a job, or Homeowners. Helpers search as their audience, so a locked note never comes back from the search. Rex gives the crew the gate code in Spanish; Juno tells the homeowner "that's not something I can see" in Chinese.
3. Access ends with the job. "Close Rex" → a confirm card → tap. Open chats stop on the spot and the site brief locks back to Office only.

Along the way the helpers get work done: Rex files a damage photo to Sam's inbox and drafts an update for the homeowner that Sam approves; questions a helper can't answer go to Sam, and his answer becomes company knowledge (and feedback to HydraDB).

Approve one card	Send the links	Talk, full screen	Crew, in Spanish	Homeowner, in Chinese	Job closed	Company brain

What's in the MVP

• One conversation for the owner. A tool-using agent (Claude) does the orchestration: create, edit, share, close, "what's new?", reply, remember a note, change a lock, ask the brain, job status. The server only provides tools and keeps the guarantees: knowledge scope is computed server-side, end times are validated, and publishing, closing, relocking and sending always need the owner's tap.
• Voice like a phone call. Full-screen voice mode on OpenAI gpt-realtime with barge-in and about 1.5 s to first audio; answers come from server-gated tools, so voice can't bypass access control. One result card per call instead of transcript noise.
• Any language. Answers follow the language people write or speak; the interface follows the phone (17 languages pre-translated, any other on demand).
• Fast, grounded answers. Sources appear in about 0.4 s, answers stream in about 2–3 s, and citations stay one tap away (collapsed by default).
• Phone first. Installable PWA for iPhone and Android, works offline for reading, queues messages until the signal is back.
• Text it, too. Homeowners and crew can text their helper over SMS (photos included) through the same access checks; a built-in SMS simulator shows it without a phone number. Twilio adapter ready (docs/SMS.md).
• Bring your Google Drive. Connect a real Drive folder through HydraDB's connector; files arrive locked to Office only until the owner relabels them (docs/DRIVE.md). Off in the public demo.
• Feels like texting a person. Short multi-bubble replies, "On it" while working, a check reaction when an action is done.
• Safe to demo publicly. Every visitor gets their own sandbox; nothing one visitor does reaches another.

Why HydraDB

CrewBrain's promise ("they only see what you allow") lives in the retrieval layer, and HydraDB made that practical in a day:

• Access lists on every source, enforced inside the search. Each helper queries as its audience, so a locked document is never retrieved, not merely hidden from the model afterwards.
• One PATCH to relock. Closing a job takes effect on the very next query, with no re-index.
• Feedback with ground truth. When the owner answers a question a helper couldn't, the answer goes back to HydraDB so retrieval improves.
• Graph context. "What's going on at 42 Oak St?" pulls related notes, issues and emails together.

How it's built

phone (PWA, voice) ─▶ server: identity → audience → lifecycle → tool gate ─▶ HydraDB (search as the audience)
        │                         │                                         └▶ Claude (owner agent, grounded answers)
        └─ gpt-realtime voice ────┘  same gated tools                          OpenAI fast model as a backup

• HydraDB is the brain: one database per company, an acl on every source derived from its lock, every helper queries as its audience (group:crewbrain:crew:<job>, group:crewbrain:household:<job>); owner answers are sent back as /feedback with ground truth; the owner's "what's going on at 42 Oak St?" uses graph context.
• Claude (Agent SDK, Sonnet and Haiku) runs the owner's agent with in-process tools and writes grounded, cited answers; document text is treated as data, never instructions.
• OpenAI gpt-realtime for speech-to-speech; a small fast model for UI translation and as a latency backup.
• Node 22 + TypeScript (Hono), plain ES-module PWA, deployed on a Mac mini behind Tailscale Funnel.

Access tests

NODE_ENV=test npx tsx --test test/*.test.ts (86 tests): create and publish, two distinct helpers, unauthorized document, tenant isolation, audience restriction, close and revocation, real expiry with a test clock, edits need approval, approval integrity, learning and escalation, prompt injection, gated write actions, answers return to the asker, crew-to-homeowner updates, voice tool gating, sandbox isolation, streaming, HydraDB feedback, the agent's tools can't publish, any-language flows.

Run it

npm ci
security add-generic-password -a crewbrain -s HYDRA_DB_API_KEY -w <key>   # also OPENAI_API_KEY, CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY
./scripts/start.sh          # http://127.0.0.1:8787/demo.html
./scripts/deploy.sh         # test-gated deploy to the mini

Without keys it runs in offline mode (local search, extractive answers, no voice).

Team

Lianghui Yi · Charlotte He · chengjiafeng857 Feng. Story and video script: Storytelling, Video Script. API: docs/API.md.