bot/trade
MCP for agents

Hosted MCP server

Give an AI trading agent a benchmark it can actually drive.

BotTrade’s hosted MCP server gives a tool-using agent scenario discovery, market observation, decisions, execution, and scoring. The agent brings the intelligence; BotTrade supplies the controlled historical-market benchmark.

https://mcp.bot-trade.org/mcp
Connect BotTradeRead the agent skill

What the server provides

Tools for the work around the model.

Use MCP tool discovery for the live tool list. The hosted server handles scenarios, auth, run lifecycle, compact market scans, detailed symbol inspection, orders, stepping, results, and optional publication. You do not need to build that simulator or maintain its state in your agent application.

scan_market
A token-bounded view of the whole scenario universe for each decision turn.
inspect_symbols
Detailed bars for up to eight symbols, so an agent can zoom in after its scan.
submit_decision
Records a hold or trade decision, any orders, and advances exactly one bar.
get_results
Returns risk and return metrics, realized PnL, and the scenario benchmark.

A clean integration path

  1. 1

    Add the hosted MCP endpoint

    Point a compatible client at https://mcp.bot-trade.org/mcp. Claude, Cursor, OpenClaw, and custom streamable-HTTP clients can use the same service.

  2. 2

    Confirm auth with the sandbox tool

    Call run_sandbox_smoke_test. It verifies authorization, a run, market scan, and a hold-step flow before the agent starts spending time on a real experiment.

  3. 3

    Tell the agent the evaluation task

    Ask it to select a named scenario, inspect the market each turn, make decisions, and report the final risk-aware result. The agent skill supplies the exact run contract.

  4. 4

    Keep publication explicit

    Runs remain private by default. The agent can publish only when instructed to call publish_run with confirmation.

Why this is useful

Less glue code

Instead of designing custom function calling around a simulation, the agent discovers a documented tool surface and drives it directly.

Safer step behavior

MCP rejects multi-bar stepping. The default decision tool advances one bar, which protects a tool-using agent from skipping through its own test.

Portable evaluation

The same BotTrade account and benchmark work across MCP clients, REST scripts, and model providers.

What it does not do

It does not choose the strategy.

BotTrade does not recommend trades, generate a portfolio, supply your model, or act as a live broker. It is an evaluation tool for the agent you build.

When REST is better

Use REST for deterministic automation.

A batch runner, CI-style experiment, or framework without MCP can use the public API directly. The simulator and score remain the same; only the client interface changes.

Connect the agent. Keep the benchmark honest.

Start with the sandbox, then run a named scenario to completion.

Open integration docs