NFL Prediction Agent

Small TypeScript agent that polls the Recall NFL competition API, gathers live context, and posts updated predictions using Vercel's AI SDK.

Requirements

• Node.js 20+ and pnpm
• Valid credentials for the Recall API and AI gateway
• Base prompt stored at prompts/game_prediction.md

Environment

Create .env with the following variables (values shown are examples):

# AI configuration
AI_GATEWAY_API_KEY=sk-***
AI_GATEWAY_BASE_URL=https://ai-gateway.vercel.sh/v1 # optional override
AI_MODEL=openai/gpt-4o-mini # optional override

# Recall API
RECALL_AGENT_API_KEY=your_recall_agent_api_key
RECALL_BASE_URL=https://api.competitions.recall.network/api # http://localhost:3000/api for local development
COMPETITION_ID=your_competition_id # or pass via --competition-id
AGENT_ID=your_agent_id # optional label for GET /predictions filtering

# General
LOG_LEVEL=info # pino log level (trace|debug|info|warn|error|fatal)

Install & Run

pnpm install
pnpm dev -- --competition-id your_competition_id

The dev command starts the polling loop immediately after resolving competition rules; use CTRL+C to stop.

Build & Deploy

pnpm build
pnpm start -- --competition-id your_competition_id

pnpm build emits dist/agent.js via tsup so the start script can be used in production environments.

Operational Notes

• The agent re-checks active games every 5 minutes and automatically skips final games.
• Prediction updates are only posted when shouldUpdatePrediction deems the delta meaningful.
• Recent play-by-play data is fetched for in-progress games (limit 50, latest first) to feed the prompt.
• Logging uses Pino; adjust LOG_LEVEL (e.g., debug) for more verbose output and optionally pipe to pino-pretty.
• AI failures or HTTP errors are logged and skipped so the loop can continue running.

Logging locally

Pipe the agent output through pino-pretty for easier reading:

pnpm dev -- --competition-id your_competition_id | pnpm exec pino-pretty

Set LOG_LEVEL=debug (or another level) to control verbosity.