Prediction market friction tracker

Structured KYC, fee, and withdrawal data across 12 prediction-market venues, plus a change feed for when those terms move and a natural-language "which venue is best for X" recommendation tool. Built for AI agents that need to check execution friction before routing a trade — not for humans clicking around a dashboard.

Live now: https://prediction-friction-tracker-production.up.railway.app

For agents: three ways in

Plain HTTP — GET /friction/venues, GET /friction/venue/{slug}, GET /friction/changes?since=<ISO8601>. Full OpenAPI spec at /openapi.json, interactive docs at /docs, agent-readable summary at /llms.txt.
MCP — a streamable-HTTP MCP server mounted at /mcp/, exposing list_venues, get_venue, recommend_venue, and get_changes_info as tools. Point any MCP-compatible client at the /mcp/ URL directly.
x402 — /friction/changes (the time-sensitive change feed) is gated by the x402 protocol: unpaid requests get a real 402 with a payment challenge; pay $0.05 USDC on Base mainnet and retry to get the data. This is live, verified end-to-end with real money, not a demo — see api.py for the CDP-facilitator wiring.

Endpoints

Endpoint	Cost	What it returns
`GET /friction/venues`	Free	Current snapshot of every tracked venue. Optional `?confidence=verified\|stale\|unconfirmed` filter.
`GET /friction/venue/{slug}`	Free	One venue's current fee/KYC/withdrawal/API-access data, e.g. `/friction/venue/kalshi`.
`GET /friction/changes?since=<ISO8601>`	$0.05 USDC (x402)	Append-only log of every detected change since a timestamp — the time-sensitive product.

Every record carries a confidence field: verified (spot-checked or high-confidence extraction), stale (source was unreachable, showing a cached value), or unconfirmed (LLM extraction was ambiguous). Treat unconfirmed as a lead to verify against the venue directly, not a fact to trade on.

MCP tools

list_venues(confidence) — browse/filter the current snapshot.
get_venue(slug) — one venue's full data.
recommend_venue(query) — free-text "which venue is best for lowest fees on high-frequency trading" / "I don't want to do KYC" / "API access without a wallet" style questions, reasoned over the current dataset. Says so explicitly rather than guessing when a query depends on data this tracker doesn't have (liquidity, volume, spread, uptime).
get_changes_info() — describes the paid change feed (URL, price) without reimplementing it for free.

What's here

File	What it does
`schema.py`	The data model — one `VenueFriction` record per venue, plus the `VenueRecommendation` response shape
`venues.py`	The 12 tracked venues and their policy-page source URLs
`scrapers/base.py`	Fetches and cleans raw policy-page text (requests + Playwright fallback for JS-rendered pages)
`extract.py`	Sends raw page text to Claude, gets back a structured `VenueFriction` matching the schema
`recommend.py`	Natural-language venue recommendation — reasons over the current dataset via Claude, structured-output constrained to real venue slugs
`reference_check.py`	Cross-checks scraped fees against a third-party source (TradeBlock), flags divergences for human review
`storage.py`	SQLite: `current_state` table (latest snapshot) + `change_log` table (append-only, the paid product)
`diff.py`	Compares a fresh extraction to the stored one, writes a `ChangeEvent` for anything that moved
`run_pipeline.py`	Orchestrates fetch → extract → diff → store, for all 12 venues
`api.py`	FastAPI server — the HTTP endpoints above, x402 payment gating, MCP mount
`mcp_server.py`	The MCP tool definitions, mounted into `api.py` at `/mcp/`
`test_pipeline_logic.py`, `test_recommend_logic.py`	Pure-logic tests, no API key needed

Running it locally

pip install -r requirements.txt
python test_pipeline_logic.py      # no API key needed, proves storage/diff logic
python test_recommend_logic.py     # no API key needed, proves recommendation logic
export ANTHROPIC_API_KEY=sk-ant-...
python run_pipeline.py             # real scrape -> extract -> store, all 12 venues
uvicorn api:app --reload           # boots the API on localhost:8000

Then in another terminal:

curl http://localhost:8000/friction/venues
curl http://localhost:8000/friction/venue/kalshi
curl "http://localhost:8000/friction/changes?since=2020-01-01T00:00:00Z"

See CLAUDE.md for environment quirks, Claude API usage notes, and the data-quality/confidence model in more depth.

Informational only — not financial or legal advice. Verify directly with the venue before acting on any figure returned here.

prediction-friction-tracker

Install to Claude Code

README