Can Claude host and run a trading strategy?

Yes — through Pancake's MCP server. Claude builds the strategy spec, validates it with a verified backtest, and deploys it with create_paper_deployment; Pancake hosts the strategy and runs it against live Polymarket data with simulated fills. Claude authors and supervises; the platform owns the runtime.

A Claude session ends; a trading strategy must not. The division of labor that resolves this: Claude writes the strategy, Pancake runs, versions, and hosts it. The strategy outlives the conversation that created it.

The flow from inside a Claude session: connect the Pancake MCP server, find evidence with search_datasets, validate the strategy with run_evidence_backtest — producing a verified result at a permanent version URL — then call create_paper_deployment with that strategy version. From that point the strategy runs on Pancake's infrastructure, not in the chat.

In any later session, Claude can check on the deployment with get_paper_deployments: open positions, simulated fills, and running P&L, all rendered from the append-only event ledger. set_paper_deployment_state pauses, resumes, or stops it. This works identically from Claude Desktop, Claude Code, or any other MCP-capable client — the surface is standard MCP tools.

Scope, honestly: deployments today are paper — live Polymarket data, simulated fills, a public ledger-backed track record. Live venue execution is a v2-roadmap capability. The point of the architecture is that a strategy Claude built carries verifiable evidence — a reproducible backtest result and an auditable forward record — rather than a chat transcript's say-so.