Crossdesk

A dark pool for business problems. Companies describe their services publicly and their problems privately. Two AI agents negotiate behind a wall — the vendor's agent is a separate model call that is never handed the buyer's problem, and a deterministic guard rejects any transcript that carries it across — and people are introduced only once the machines agree, with a briefing already written.

Hackathon, 18–19 September 2026. Live at https://crossdeskmatch.vercel.app.

Document	What's in it
PLAN.md	Scope, demo path, architecture, timeline
docs/RESEARCH.md	Market research, competitors, answers for the judges
docs/JUDGING.md	The scoring criteria, where we stand, what moves each score — read before deciding what to build next
docs/FRONTEND.md	The verified backend contract, screen by screen
TEAM.md	Who owns which files, how we stay in sync
CLAUDE.md	Rules every Claude Code session must follow

Getting started

git clone https://github.com/Dennel04/b2b-match.git
cd b2b-match
npm install
cp .env.example .env.local   # fill in the keys, ask the backend owner
npm run dev

Supabase

1. Create a project at supabase.com.
2. SQL Editor → run every file in supabase/migrations/ in order (0001, then 0002, …). They are forward-only; a project that has run 0001 only needs the newer ones. 0006 wants npm run backfill:candidates after it, and 0007 is the one that gives each service its own terms — without it /services is empty and publishing one fails.
3. Copy the keys from Settings → API into .env.local.

Commands

npm run dev     # development server
npm run check   # route typegen + typecheck + every *.check.ts — run before every push
npm run seed    # 18 seeded companies with profiles, terms and problems

npm run seed needs SEED_OWNER_ID in .env.local — the id of any registered user.

The two interview prompts have a bench each. They call the real model, so they cost money and are deliberately out of check:

npx tsx --env-file=.env.local scripts/bench.ts          # the buying side
npx tsx --env-file=.env.local scripts/bench-service.ts  # the selling side
npx tsx --env-file=.env.local scripts/interviews.ts     # read real interviews back out of the database

Working in parallel

Everyone runs their own Claude Code session; the shared context lives in git, not in one session. A hook pulls teammates' changes before every prompt and tells the model to re-read any shared file that changed. See TEAM.md.