What is MCP backtesting?

MCP backtesting is running a strategy backtest from inside an AI agent via the Model Context Protocol (MCP) — the agent calls structured tools (search_datasets, register_evidence_dataset, run_evidence_backtest, get_backtest_result) and receives a verifiable receipt without leaving its session.

The Model Context Protocol (MCP) is an open protocol that lets AI agents call structured tools exposed by external services. Pancake publishes an MCP endpoint at mcp.usepancake.com that exposes seven tools: search_strategies, search_datasets, get_strategy, register_evidence_dataset, run_evidence_backtest, get_backtest_result, and materialize_evidence_dataset. Pancake is hosting infrastructure for AI-built trading strategies — MCP is the interface through which any MCP-capable agent hosts and validates strategies on the platform.

In an MCP backtesting session, the agent assembles a strategy spec in the Pancake evidence format, finds or uploads matching evidence rows, submits the run, and receives a receipt short_id. The full receipt — including all metrics and the verification boundary — is then available as Markdown at /r/<short_id>.md, which the agent can read directly in its context window. Backtest is the on-ramp; the receipt is the artifact the strategy carries toward live execution (a v2-roadmap capability).

The key advantage over conventional API integrations is that the agent never leaves its session: all steps are tool calls, results flow back as text, and the receipt URL is a permanent citation target the agent can include in its output to the user.

Pancake uses Streamable HTTP transport per the MCP spec. OAuth 2.0 handles authorization with six scopes. The first tool call triggers a consent screen; subsequent calls are authenticated automatically.