Quickstart

Connect any MCP-capable agent to Pancake in under five minutes.

Connect Pancake to your LLM

Pancake exposes a Model Context Protocol (MCP) server. Any MCP-capable agent can connect to it with a single endpoint URL — no SDK required.

Add the following to your agent's MCP configuration:

{
  "mcpServers": {
    "pancake": {
      "url": "https://mcp.usepancake.com/api/mcp",
      "transport": "http"
    }
  }
}

The server uses Streamable HTTP transport. If your agent host requires an explicit transport field, use "http". OAuth authorization happens on first tool call — the server returns a redirect to the Pancake consent screen if no token is present.

Authorize via OAuth

Pancake uses OAuth 2.0 to authorize agent access. When your agent calls any tool for the first time without a valid token, Pancake returns a consent redirect. Approve the following six scopes:

Scope	What it grants
`read:markets`	Browse canonical market pools and metadata.
`read:strategies`	Read strategy specs you own or that are public.
`read:datasets`	Browse and fetch EvidenceDatasets from the canonical pool.
`write:strategies`	Create and update strategy specs in your account.
`read:results`	Fetch backtest receipts and their computed metrics.
`run:backtests`	Submit a strategy + evidence pair to the engine and receive a receipt `/r/<short_id>`.

All six scopes are bundled in the default consent screen. You do not need to request them individually. Tokens are stored in your agent host and refreshed automatically.

Your first backtest in one sentence

Tell your agent what you want to test — in plain English. Pancake handles the rest:

"Backtest a momentum strategy on Polymarket GPU markets over the last 90 days."

Under the hood, the agent invokes three Pancake tools in sequence:

search_datasets — finds a matching EvidenceDataset in the canonical pool (or prompts you to register one).
register_evidence_dataset — if no pool match exists, uploads your evidence rows and validates them for structural invariants (schema, lookahead, monotonicity, range).
run_evidence_backtest — submits the strategy spec and evidence to the batter engine. Returns a canonical receipt URL: /r/<short_id>.

The full round-trip typically completes in under 10 seconds. The agent surfaces the receipt URL and a plain-English summary of the key metrics.

What you get back

Every backtest produces an immutable receipt at usepancake.com/r/<short_id>. Receipts carry three classes of information:

Verified metrics

Computed by the batter engine from your evidence rows. Byte-stable across runs — confirmed by a SHA-256 result_hash in the receipt envelope. Includes total return, CAGR, Sharpe, Sortino, max drawdown, win rate, and bootstrap CI on each.

Verification boundary

Every receipt declares what Pancake verified structurally, what your agent supplied as evidence (and Pancake did not independently source), and what risk factors were not modeled. No hidden assumptions. See Methodology for the full taxonomy.

Engine stamp

Each receipt carries engine_version so any reader can reproduce the result by running batter at the stated version against the cited dataset. The receipt is the proof.

Receipts are available as HTML (/r/<short_id>), Markdown (/r/<short_id>.md), and JSON (/api/v1/public/results/<short_id>).