wc26bot

A working example of an eve agent with tools, schedules, channels, evals, and a custom web UI.

The football use case gives each part something real to do. The agent reads current fixtures and results, compares Polymarket and Kalshi prices, collects source-backed tournament facts, and settles fictitious exact-score predictions. The same agent runs on the web, Slack, and X.

What to try

• Change the agent's voice and grounding rules in agent/instructions.md, then add channel-specific behavior in agent/instructions/surface.ts.
• Add a typed tool under agent/tools. The examples cover live data, durable state, and custom UI results.
• Run background agent work from agent/schedules, including source collection before kickoff and prediction settlement after full time.
• Expose one agent through the web, Slack, and X with agent/channels.
• Test conversation behavior with live evals and deterministic domain logic with test.

Run locally

You need Node.js 24 or newer, pnpm, and a Vercel AI Gateway credential.

git clone https://github.com/vercel-labs/wc26cup.git
cd wc26cup
pnpm install

Create .env.local:

AI_GATEWAY_API_KEY=your_ai_gateway_key
BLOB_READ_WRITE_TOKEN=your_vercel_blob_token

The Blob token enables the fact memory, predictions, and X channel state. You can use VERCEL_OIDC_TOKEN instead of an AI Gateway key; pnpm exec eve link pulls it from a linked Vercel project.

pnpm dev

Open http://localhost:3000, or run pnpm dev:eve for eve's terminal UI.

Useful checks:

pnpm typecheck
pnpm test:unit
pnpm exec eve eval --strict
pnpm build:eve

Live evals need both credentials shown above. Unit tests and typechecking need neither. Run pnpm exec eve deploy to deploy a linked Vercel project.

Reference

• eve
• AI SDK
• AI Elements
• Chat SDK
• Next.js
• Streamdown
• Satori
• Vercel Blob SDK
• Vercel Connect SDK
• Geist

Market prices are time-stamped snapshots, not forecasts or betting advice. Predictions are fictitious and involve no money.