Polymarket Wallet Intelligence

An MCP server and REST API that classifies Polymarket wallets as human or bot, scores their trading edge from 0–10, and streams their current open positions. Built for AI agents on copy-trading and signal-following stacks.

# Use it from any MCP client (Claude Desktop, Cursor, etc.)
pip install polymarket-intel-mcp
polymarket-intel-mcp

# Or call the hosted REST API directly
curl https://polymarket-intel-production.up.railway.app/wallet/0xf1528f12e645462c344799b62b1b421a6a4c64aa

How this fits with other Polymarket MCP servers

Status

Latest: v1.2 (May 2026) — classifier improved to distinguish active human grinders from HFT bots. See release notes and v1.3 backlog issue for what's next.

A daily snapshot job runs at 08:00 UTC and re-scores the top 50 leaderboard wallets, building a historical dataset of classification stability over time.

There are several MCP servers covering Polymarket, each at a different layer:

These complement each other. A copy-trading agent would use polymarket-intel to filter wallets worth following, then graph-polymarket-mcp to read the markets those wallets are betting on, then whitmorelabs/polymarket-mcp to size its own entries.

What it answers

• "Is this trader a human or a bot?" — score_polymarket_wallet(wallet_address) → returns classification ∈ {human, bot, insufficient_data} plus a confidence score and reason codes.

• "Do they actually have an edge?" — edge_score from 0–10, gated on net realised PnL so distributed-but-losing wallets don't get false positives.

• "What are they betting on right now?" — get_open_positions(wallet_address) returns live positions sorted by size, refreshed every 30s.

• "How has their edge changed over time?" — /wallet/{address}/history returns the score time series from the daily snapshots.

Why this exists

The Polymarket leaderboard is misleading. It includes unrealised PnL marked-to-current-price, so the names at the top are dominated by bots running structural arb plus a few wallets sitting on huge open positions that may never resolve in their favour. Agents that copy-trade naively from the leaderboard get burned.

This service runs every leaderboard wallet through behavioural fingerprinting (focus ratio, holding period, timing regularity, category concentration) plus PnL reconstruction from raw activity, and only surfaces traders that look like genuine humans with a real edge.

The dataset grows more valuable over time — every day the snapshot job runs, historical signals accumulate. Wallets that have been consistently above edge 7 for 90 days are a stronger signal than any single point-in-time score.

Distributed as both a REST API and an MCP server

Architecture

┌──────────────────────────────────────────────┐
│  core/                                       │
│    client.py    — Polymarket data API client │
│    signals.py   — pure signal calculators    │
│    scorer.py    — classifier + edge score    │
│    models.py    — Pydantic response schemas  │
├──────────────────────────────────────────────┤
│  db/                                         │
│    schema.sql   — Postgres tables + indexes  │
│    repository.py — Repository protocol +     │
│                    InMemoryRepository        │
│    supabase_repo.py — Supabase impl          │
│    converters.py — ScoreResult ↔ records     │
├──────────────────────────────────────────────┤
│  api/main.py    — FastAPI HTTP server        │
│  mcp_server/    — MCP server (stdio)         │
│  scripts/                                    │
│    analyze_wallet.py — CLI                   │
│    snapshot_job.py   — daily cron entry      │
│  tests/                                      │
└──────────────────────────────────────────────┘

Core has no idea persistence exists. The API and snapshot job depend on the Repository protocol — Supabase in production, in-memory in tests and when env vars are unset. This is what makes the suite run without a database and what lets you swap Supabase for Neon, RDS, or anything else later by adding one file.

Quickstart

git clone <repo> && cd polymarket-intel
python -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
pytest                             # 19 tests, all green

CLI

python scripts/analyze_wallet.py phonesculptor
python scripts/analyze_wallet.py 0xf1528f12e645462c344799b62b1b421a6a4c64aa --json

REST API

uvicorn api.main:app --reload --port 8000
open http://localhost:8000/docs

The API is split into a slow tier (cached aggressively, cheap, ideal for one-off discovery) and a fast tier (short cache, ideal for live copy-trading agents). The split exists because the underlying data has different freshness needs — a wallet's classification doesn't change minute-to-minute, but their open positions do.

Why 30s on positions and not faster? Polygon block time is ~2s and Polymarket's activity index lags a few seconds. Polling below 10s gets you no fresher data, just rate-limit errors. 30s is the sweet spot for cost/freshness/upstream-friendliness.

Why debounced DB writes? A trading agent may hit /wallet/{address} thousands of times an hour. Writing a row per call would bloat history with near-duplicate snapshots. The score endpoint persists at most once per wallet per hour. The daily snapshot job guarantees coverage of the top 50 regardless of API traffic.

MCP server (Claude Desktop, Cursor, Continue)

python mcp_server/server.py

Then drop mcp_server/claude_desktop_config.example.json into your Claude Desktop config and edit the absolute path.

The server exposes four tools:

• score_polymarket_wallet(wallet_address) — full score

• score_polymarket_user(username) — lookup by display name

• get_polymarket_leaderboard(limit) — raw leaderboard

• get_open_positions(wallet_address) — fast snapshot of live bets

Scoring methodology

Bot triggers (any one fires → bot)

Soft signals stack: crypto-market-maker pattern, > 200 trades/day, etc.

Edge score (0–10) for humans

Hard gate: net realised PnL ≤ 0  →  capped at 2.0
Hard gate: < 10 winning markets  →  capped at 3.0

35%  PnL magnitude (log-scaled)
25%  win rate (capped at 70%)
15%  PnL distribution (penalises top-1 concentration)
15%  sample size (winning markets, capped at 50)
10%  win/loss ratio (capped at 3x)

Net PnL is the hard gate so wallets like neutralwave23 — many distributed tiny wins masking $375k of losses — are correctly flagged as poor.

PnL reconstruction

money_in   = sum(BUY usdcSize per conditionId)
money_out  = sum(SELL usdcSize) + sum(REDEEM usdcSize)
pnl        = money_out - money_in

status:
  REDEEM exists                                    → won
  SELL exists, no REDEEM                           → exited
  no SELL, no REDEEM, last trade > 7 days old      → lost
  no SELL, no REDEEM, last trade within 7 days     → open

Why activity rather than the positions endpoint: positions vanish from the API after redeem, so any naive analysis using /positions undercounts wins. Always reconstruct from /activity?type=TRADE + /activity?type=REDEEM (separate calls — comma-joined types return 400).

Persistence (Supabase)

The historical dataset is the moat. Every day the snapshot job pulls the leaderboard, scores the top N wallets, and persists three things: the score itself (wallet_scores), the wallet's open positions at that moment (open_position_snapshots), and the leaderboard as it stood (leaderboard_snapshots). After 90 days you can answer questions no one else can: "who has been consistently above edge 7 for the last quarter?", "which wallets just entered the top 50?", "show me everyone who held YES on this market three days before resolution."

Setup

# 1. Create a Supabase project, get the URL and service_role key
cp .env.example .env  # fill in SUPABASE_URL and SUPABASE_KEY

# 2. Apply the schema (Supabase dashboard → SQL editor → paste db/schema.sql → run)
#    Or via psql:
#    psql "$DATABASE_URL" -f db/schema.sql

# 3. Run the snapshot job once to verify it writes:
python scripts/snapshot_job.py --top 10

If SUPABASE_URL and SUPABASE_KEY are unset, both the API and the snapshot job fall back to an in-memory repository — the suite still passes, the API still serves live scoring, but history endpoints will be empty until you wire up Supabase.

Daily snapshot job

Schedule python scripts/snapshot_job.py --top 50 daily (Railway cron, GitHub Actions, or Supabase pg_cron triggering an edge function — your call). The job is idempotent: running twice creates two snapshots, which is fine — history queries pick the closest one.

python scripts/snapshot_job.py --top 50              # production
python scripts/snapshot_job.py --top 5  --dry-run    # local testing, no writes

Each run records an audit row in snapshot_runs with start/finish times, wallets scored, and error count.

Schema overview

Two views (latest_wallet_scores, latest_leaderboard) make the common "what's current" queries cheap.

Repository pattern

db/repository.py defines a Repository protocol. Two implementations:

• InMemoryRepository — thread-safe, lossy across restarts. Used in tests and as the dev-mode fallback.

• SupabaseRepository — production. Wraps the supabase-py client.

The API and snapshot job depend only on the protocol. To swap Supabase for Neon or self-hosted Postgres, write one new class implementing the same six method signatures.

Pricing dimensions

The endpoint split was designed so each tier maps cleanly to a billing model. Suggested ranges:

The streaming endpoint is the one serious copy-trading bots will actually pay for, but it requires a continuous-fetch worker on our side — leaving it for v2 once we have signal that the per-call business works.

Deploy

Railway

Push the repo, point at it. railway.toml handles the rest.

Render / Heroku-style

Procfile is in place.

Caching

api/cache.py is a thread-safe in-memory TTL cache with the same interface as a Redis client. For multi-worker production, swap the singleton for redis.Redis() in one file. TTLs:

• wallet score: 1h

• open positions only: 5m

• leaderboard: 30m

• verified leaderboard: 1h

Testing

pytest -v

Synthetic fixtures in tests/fixtures.py mimic the three real wallet patterns from the research phase (phonesculptor MLB human, gabigol HFT bot, neutralwave23 tilt loser) plus a low-data newbie. Tests run against fixtures only — no live API calls — so the suite is deterministic and CI-safe.

Distribution roadmap

1. Now — REST API on Railway, Supabase for daily snapshot persistence

2. Next — Publish to MCP Hub, Replit Agent Market, awesome-mcp-servers

3. Later — x402 micropayments per call (USDC), historical query endpoints (the moat: every day we run, the dataset grows)

Verified wallet examples

These are the personas the test fixtures target. Live numbers will differ as activity changes: