A finance MCP server for allocation models and research

A finance MCP server gives an AI assistant tools it can call while researching investments. ENSEMBLE provides a hosted server for the complete allocation-model workflow: find assets, define a rule, test it, inspect the result, and maintain its allocations. An allocation describes how investments are divided among assets or other models.

The assistant works with the same saved models available in the web app, API, and CLI. A conversation can end while the model, its code, inputs, builds, and runs remain available for later work. ENSEMBLE is research software; the outputs are hypothetical research and target allocations.

Before connecting

Create an ENSEMBLE account and check the MCP connection guide for a supported client. The client must support a remote Streamable HTTP server and either OAuth or a bearer API key. A URL alone does not establish compatibility: client plans, administrator permissions, and support for write actions can differ.

The server URL is https://api.ensemble.markets/mcp. Prefer OAuth where supported. For a headless client, create an ENSEMBLE API key and configure it as a bearer token using the client's documented settings. Keep keys out of prompts and shared configuration files.

Start with $100 in credit, valid for 7 days, with no card required. Builds and runs are metered. See pricing for current plans and usage rates. Creating, rebuilding, testing, and refreshing models can consume credit. Connecting an agent does not authorize it to place brokerage orders; ENSEMBLE does not execute trades.

Data and execution are both available

The tools expose asset metadata and historical prices as well as model operations. list_assets and get_asset with optional include=["history"] help establish which instruments and dates are available. Coverage and metadata vary by asset; a missing field remains unknown.

The runtime uses daily data. Inspect each asset's coverage, the model's selected window, and its execution assumptions. Do not infer a complete historical universe, universal delisting coverage, or a fixed number of years from the existence of a ticker. Backfills can extend history using an explicit proxy and transformation; those periods are synthetic, not observed fund performance.

Models run in an isolated environment with resource, time, and composition limits. Supplied Python must satisfy the current model runtime contract. The methodology explains how to interpret the research record and its limitations.

Tools for the workflow

Discover the current tool schemas from the connected server. These names describe the main stages; the schemas supply their accepted arguments.

Builds may continue after the create call returns. Inspect the model's in-progress run and status until it finishes or reports an error. Completion time depends on the work and selected effort; there is no fixed build-time promise. A queued run is not a successful result.

Use parameter tests when only declared inputs change. For strategy-code changes, use a separate private model when preserving the approved implementation; revision tools are not currently enabled in the plugin. A new prompt can produce different code, so rebuilding is not equivalent to varying one parameter.

A useful first request

Find SPY and BIL in the catalog and report their available history. Describe a monthly model that holds SPY above its 200-day moving average and BIL otherwise. Explain the proposed timing and cost assumptions before building. After building, inspect the code, the selected dates, and the latest successful run.

This is an example request, not a recorded conversation or a performance result. The agent must inspect the tool responses and explain any difference between the requested rule and the implementation. The Claude walkthrough demonstrates that habit in a specific client.

Where the platform stops

The server does not provide real-time quotes, an intraday or options backtesting engine, a standalone financial-planning simulator, or brokerage execution. Supported broker-format files are outputs for the user's execution workflow; they do not establish broker tradability or acceptance.

Backtests do not predict future returns. Changed market data, optimizer code, or model dependencies can affect later results even when the top-level script is unchanged. Read the stored evidence rather than treating an assistant's summary as verification.