Use case · Agents

An AI agent trading desk

polybot's MCP server turns any compatible agent into a research analyst with a real trading desk behind it — one that can read the book, assess strategies, and draft trades you approve.

What the agent can do

• Introspect. list_markets, get_positions, get_orders and get_strategy_code let it explain exactly what is happening.
• Assess. analyze_strategy and compare_strategies query the DuckDB analytics store over a window and return P&L, Sharpe, fill rate and per-pair breakdowns.
• Propose. suggest_strategy_improvements is grounded in the strategy's code plus its performance report — not a blind prompt.
• Draft, never fill. propose_config_change and order proposals are written to disk and wait for your approval.

A worked session

Ask Claude to review the last two weeks, pick the worst strategy and suggest one concrete change. It calls the assessment tools, reads the code, spots that low-z entries are losing, and proposes raising the entry threshold — then writes a pending proposal. You inspect it with polybot proposal show and apply it. The full annotated walkthrough is in the agent loop guide.

Why this is safe

The agent is a first-class principal with tools, memory and audit — not a script with an API key. Every tool call is logged with actor and timestamp, and execution stays gated behind shadow mode and human approval. See risk controls and how it works.

FAQ

Which agents can drive polybot?

Any MCP-compatible client — Claude Desktop, custom agents built on the MCP SDK, or IDE extensions that speak MCP. polybot exposes its tools over stdio or HTTP.

Can the agent place a live trade on its own?

No. Execution is approval-gated. An agent can propose a config change or draft an order, but polybot writes a proposal and waits for you to approve it at the CLI before anything goes live.