manifoldbt-mcp

Releasing to PyPI is fully automated via GitHub Actions + PyPI Trusted Publishing — just tag vX.Y.Z and push.

Model Context Protocol (MCP) server for manifoldbt — the Rust-powered backtesting engine for quantitative research.

Expose the full manifoldbt backtesting engine to any MCP-compatible client (Claude Desktop, Claude Code, Cursor, Devin, VS Code, Windsurf, …). Let an LLM author strategies, run backtests, sweep parameters, and generate tearsheets — without the user writing a line of Python.

This package is not affiliated with the manifoldbt authors; it is a thin adapter that imports the published manifoldbt wheel and wraps its Python API behind an MCP server. All heavy lifting is done by the underlying Rust engine.

Install

pip install manifoldbt-mcp

This pulls in manifoldbt and the reference mcp Python SDK as dependencies. Python 3.10+ is required.

Optional extras:

pip install "manifoldbt-mcp[plot]"   # plotly stack, for plot_tearsheet
pip install "manifoldbt-mcp[dev]"    # pytest, ruff, plotly stack

Run

manifoldbt-mcp                          # stdio transport (default)
manifoldbt-mcp --http --port 8765       # streamable-HTTP transport
manifoldbt-mcp --http --host 0.0.0.0    # remote deployment

Client configuration

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "manifoldbt": {
      "command": "manifoldbt-mcp"
    }
  }
}

Claude Code

claude mcp add manifoldbt -- manifoldbt-mcp

Cursor / Windsurf

{
  "mcpServers": {
    "manifoldbt": {
      "command": "manifoldbt-mcp"
    }
  }
}

VS Code (Continue, Cline, …)

Any MCP-over-stdio client will work — just point it at the manifoldbt-mcp binary that pip installed on your PATH.

Remote (HTTP)

manifoldbt-mcp --http --host 0.0.0.0 --port 8765

Then connect your MCP client to http://<host>:8765/mcp.

What is exposed

20 tools

3 resources + 1 template

2 prompts

• write_strategy(description, universe="BTC-USDT:perp") — scaffold a new strategy from a natural-language description.
• analyze_result(metrics_json) — turn a run_backtest metrics payload into a research note.

Ergonomics

Every tool is JSON-friendly so LLM clients don't have to think about types:

• Dates — ISO strings: "2022-01-01" or "2022-01-01 09:30:00".
• Intervals — shorthand: "1h", "5m", "15s", "1d".
• Fees — preset names: "binance_perps", "binance_spot", "zero"; or a dict with individual overrides.
• Slippage — {"kind": "fixed_bps", "bps": 2}, {"kind": "volume_impact", "impact_coeff": 0.1, "exponent": 1.5}, etc.

Writing a strategy from a client

Strategies are authored as Python DSL snippets executed in a namespace that pre-imports every manifoldbt symbol — so the LLM never has to write import statements:

fast = ema(close, 12)
slow = ema(close, 26)

strategy = (
    Strategy.create("ema_cross")
    .signal("fast", fast)
    .signal("slow", slow)
    .size(when(fast > slow, lit(0.5), lit(0.0)))
    .stop_loss(pct=3.0)
)

Or, if the client prefers, send the already-serialised StrategyDef JSON through strategy_json on any tool.

Example run_backtest payload

{
  "strategy_code": "fast = ema(close, 12)\nslow = ema(close, 26)\nstrategy = Strategy.create('ema_cross').signal('fast', fast).signal('slow', slow).size(when(fast > slow, lit(0.5), lit(0.0))).stop_loss(pct=3.0)",
  "config": {
    "universe": {"binance": ["BTC-USDT:perp"]},
    "start": "2022-01-01",
    "end": "2024-01-01",
    "bar_interval": "1h",
    "initial_capital": 10000,
    "fees": "binance_perps",
    "slippage": {"kind": "fixed_bps", "bps": 2},
    "warmup_bars": 60
  },
  "store": {
    "data_root": "data",
    "metadata_db": "metadata/metadata.sqlite"
  }
}

Typical LLM session

1. list_indicators → discover available indicators.
2. Use the write_strategy prompt → draft a strategy.
3. build_strategy(strategy_code=…) → validated StrategyDef + declared parameters.
4. run_backtest(…) → metrics + summary.
5. run_sweep(…) ranked by Sharpe → find the best variant.
6. plot_tearsheet(…) → HTML tearsheet report on disk.
7. Use the analyze_result prompt → research note.

Security

• The server runs with the permissions of the launching process.
• strategy_code is executed in-process by the Python interpreter — there is no sandbox. The DSL namespace is convenient (pre-imported manifoldbt symbols), but nothing prevents arbitrary Python from running. Only use this server with trusted clients, or run it isolated (dedicated venv, container, restricted user).
• No inbound network unless you explicitly enable --http.

Development

git clone https://github.com/ClementG91/manifoldbt-mcp
cd manifoldbt-mcp
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest
ruff check .

Related

• Upstream engine: https://github.com/Jimmy7892/manifoldbt
• Model Context Protocol: https://modelcontextprotocol.io
• Reference Python SDK: https://github.com/modelcontextprotocol/python-sdk

License

MIT — see LICENSE.