— Predictions Intelligence Service / N° 002
The predictions
knowledge base
for AI agents.
Evens is the data layer AI agents call when they need real-time prediction-market intelligence. Consensus prices across Kalshi and Polymarket, cross-venue arbitrage, whale signals, and topic-level synthesis. One MCP install. One source of truth.
Why now
Two market
conditions just changed.
AI agents are buying real-time data.
Tavily, Perplexity API, Exa, Linkup, Serper — each crossed $10M ARR selling "data for AI agents." Tavily alone runs ~100M monthly requests. None of them touch prediction markets.
MCP became the integration surface.
Claude Code, Cursor, Windsurf, ChatGPT Apps all consume MCP. Forrester predicts 30%+ of enterprise vendors ship MCPs in 2026. The only listed predictions-market MCP today has 29 generic tools and no positioning.
“Tavily owns horizontal web-search for agents. Evens owns the predictions vertical.”
Built & running
The infrastructure
already exists.
We've spent the past year building the canonical data layer underneath the rebrand. ~80% of the work for the agent product is repackaging what's already in production.
Canonical data model
75+ models normalising events, markets, references, prices across 4 venues.
Internal MCP server
70+ tools in production for sync, remap, template debugging.
REST + GraphQL + WebSocket
OpenAPI 3.1 documented. Live price feeds. Already used by paying customers.
Arbitrage scanner
Cross-venue spread detection with worked-payoff playbooks.
Whale tracker
Large-position detection across providers with size estimation.
Webhook + alert engine
Threshold rules with retry logic; ready to push signals to subscribers.
The wedge
One feature.
One install.
One brand line.
We're shipping a public, registered MCP server tuned for agent reasoning — not internal ops. Curated 10-tool surface, with one synthesis tool nobody else builds.
Read & signal tools
search_markets, find_arbitrage, recent_whale_trades, price_history. Reuses scanners already in production.
prediction_brief
Single-call synthesis. Topic in, full intelligence packet out. Today an agent needs ~5-7 separate calls to assemble this.
subscribe_alert
Agent registers a webhook for a price / arb / whale condition. Closes the loop — we push, they listen.
Defensibility
Four moats.
Compounding.
Distribution
Registered in the Anthropic MCP registry on day one. Every Claude Code, Cursor, and Windsurf user discovers Evens through claude mcp browse. First-mover lock-in on the install is structural.
Data flywheel
More agents calling = more usage signal on what queries matter = better-tuned synthesis = more agents calling. The same loop Tavily ran from $0 to $10M ARR in 18 months.
Multi-source synthesis
Polymarket / Kalshi will eventually ship single-venue MCPs. They can't ship cross-venue arbitrage, whale signals, or consensus pricing — that requires aggregator infrastructure.
Switching cost
Once an agent's prompts depend on prediction_brief's response shape, replacing the source means rewriting the agent. The response contract is itself the lock-in.
“What if Polymarket / Kalshi ship their own MCP?” They will — for their venue. Our moat is the multi-source layer they can't replicate.
Numbers
TAM today.
Market in 2028.
Prediction-market data tooling. ~5K serious traders × ~$1.5K ACV. The current SaaS book.
Agent-data infrastructure. Tavily / Exa / Perplexity API are the early proof points. Capturing 1-2% × predictions-relevance share.
Sprint exit → Month 3
| Metric | Sprint exit | Month 1 | Month 3 |
|---|---|---|---|
| MCP installs (registry) | 50 | 500 | 5,000 |
| Weekly active calling agents | 20 | 200 | 2,000 |
| Tool calls / week | 5K | 100K | 1M |
| Paying customers | 5 | 30 | 200 |
| MRR | $300 | $3K | $20K |
Pricing — mirrors Tavily / Perplexity API
100 calls/day. Read tools only.
10K/day. prediction_brief unlocked.
Unlimited. Whale signals + alerts.
SSO, custom prompts, private deploy.
30 customers at $99/mo = $35.6K ARR. Covers infra and the founder. Default-alive without a Series A unlock event.
Let's talk.
Two-week sprint to public launch. Already-built data layer underneath. Distribution-moat playbook with a six-month head start before any incumbent ships their own MCP.