Ask your assistant whether a Polymarket wallet is worth copying.

Claude

Settings → Connectors → Add custom connector, and paste the URL above. Available on paid plans, on the web and desktop apps.

Claude Code

claude mcp add --transport http copygrade https://copygrade.com/api/mcp

ChatGPT

Settings → Connectors → Advanced → Developer mode, then add the URL as a custom connector. Paid plans only, and the toggle is buried — expect this to be the fiddliest of the three.

Cursor, or any agent framework

Register it as a remote MCP server over Streamable HTTP. No credentials, and no session handshake to keep alive — every call is one request in, one response out.

Prefer plain HTTP? The same read is one request, no key:

Takes one argument — address, a wallet's 0x… as shown in its Polymarket profile URL. Read-only and idempotent: it fetches a published verdict and changes nothing. An address CopyGrade has not indexed yet returns a not-found the model can relay, pointing at the free live grade rather than inventing a score.

{
  "address": "0x4a8c…f21b",
  "handle": "someTrader",
  "score": 82,
  "verdict": "pos",
  "label": "Copy candidate",
  "subScores": {
    "authenticity": 84,
    "sharpe": 79,
    "drawdown": 71,
    "consistency": 88,
    "farming": 95
  },
  "farming": { "level": "clean", "signals": 0 },
  "edge90d": 18.4,
  "edgeReal": 12.1
}

edge90d is the wallet's headline 90-day edge; edgeReal is the same edge after modelled fees and copy latency — the number that matters to a follower rather than to the trader. The five sub-scores and the farming level are the same values the public verdict page shows, and the payload is identical to GET /api/v1/wallets/{address}, so an agent that learns one surface can read the other.

list_copyable_wallets{ limit? }

The public shortlist — farming-clean, CopyGrade Score 65 or above, at least one fill in the last 30 days — best score first, with the graded total as the honest denominator. The same list as the best-traders page. It names its bar in the answer, and it is not the Scout's stricter four-part copyable bar; a shortlist is a place to start due diligence, never a pick.

screen_wallets{ verdict?, farming?, minScore?, category?, active30d?, limit? }

Rank the graded set under a small, closed filter set — verdict band, farming flag, minimum score, market category, active this month — up to 50 rows, with the size of the whole match. An unknown filter is refused, never ignored, so a misspelt argument can't silently return the unfiltered set.

simulate_copy{ address, capital?, days?, latencyMs?, slippageBps?, botFeeBps? }

The Copy Simulator as a tool: replay a wallet's real fills scaled to your capital, charged Polymarket's published taker fees plus a bot fee, slippage and latency, and get back the net return, worst drawdown and what each cost erased. Too few fills in the window and it says so — a prior-based estimate, flagged as one.

The two list tools have plain-HTTP twins — GET /api/v1/wallets/copyable and GET /api/v1/wallets/screen — on the same budget, so a script and an agent read the same answer.

• Grade Polymarket wallet 0x… — is it worth copying, and what's the farming risk?
• Which Polymarket wallets currently clear CopyGrade's shortlist bar, and how many wallets was that drawn from?
• Screen for farming-clean Sports wallets with a CopyGrade Score of 70 or more that traded this month.
• Simulate copying wallet 0x… with $1,000 over the last 30 days at 500 ms latency — what did fees, slippage and latency cost?
• Compare these three Polymarket wallets on CopyGrade Score and post-fee edge, and tell me which has the steadiest consistency sub-score.
• Check this wallet's CopyGrade verdict before I copy it, and flag anything in the sub-scores that contradicts the headline score.
• What does CopyGrade's farming-risk level mean for this wallet, and does its edge survive fees and latency?

• Analysis only. CopyGrade never executes trades, holds funds or custodies anything, and every tool this server exposes is read-only. There is no write path to expose.
• No authentication required. The budget is per IP — 20 calls a minute and 200 a day, shared with the keyless HTTP endpoint so the two cannot be combined for a larger allowance. Over the budget you get a clear retry-after rather than a silent failure. A Pro API key sent as a Bearer header lifts a tool call onto the key's own daily quota of 1,000 calls; the handshake is free either way.
• Stateless Streamable HTTP. JSON responses only, no SSE stream and no session ids. Protocol revisions 2025-06-18 and 2024-11-05 are negotiated.
• A score is a risk opinion, not advice. The model is v1 and openly unvalidated against forward copier returns — we publish that read ourselves rather than wait to be asked. See the methodology for how the score is built and what it does not claim.