backtest360-mcp

backtest360-mcp

MCP server exposing the Backtest360 engine API as tools for AI agents.

Connect any MCP-capable AI client and drive real backtests conversationally: discover indicators, build and validate strategies, run backtests, and read the results — all against the deterministic Backtest360 engine. The server contains no AI and computes no numbers of its own; it is a thin, faithful adapter over the engine HTTP API. Your engine API key and its plan govern everything (permissions, rate limits, data access).

Two transports: a hosted HTTP endpoint at https://mcp.backtest360.com/mcp (send your key as an X-API-Key header) and local stdio (self-host — see below).

Install

pip install backtest360-mcp        # or, from a clone: pip install -e .

Requires Python 3.10+ and a Backtest360 API key. Get one free, instantly at backtest360.com/api-access — submit your email and a key (format b360_…) is issued on the spot and emailed to you; no approval needed. Authentication is API-key only. The free tier runs backtests on data you upload; fetching historical price data from the engine server-side is a paid capability.

Configuration

Everything is environment-driven:

Variable	Required	Default	Purpose
BACKTEST360_API_KEY	yes	—	Engine API key, sent as X-API-Key
BACKTEST360_ENGINE_URL	no	https://api.backtest360.com	Engine base URL
BACKTEST360_MCP_TIMEOUT	no	300	Per-request timeout (seconds)
BACKTEST360_MCP_MAX_OUTPUT_BYTES	no	100000	Hard cap on a single tool result

Connect an MCP client

Hosted (recommended)

Point your MCP client at the hosted endpoint over HTTP and send your key as an X-API-Key header:

{
  "mcpServers": {
    "backtest360": {
      "type": "streamable-http",
      "url": "https://mcp.backtest360.com/mcp",
      "headers": {
        "X-API-Key": "b360_..."
      }
    }
  }
}

Local (stdio)

Run the server yourself and let your client launch it over stdio (the common mcpServers shape):

{
  "mcpServers": {
    "backtest360": {
      "command": "backtest360-mcp",
      "env": {
        "BACKTEST360_API_KEY": "b360_..."
      }
    }
  }
}

Prefer not to put the key in a config file? Point command at a small wrapper script that exports the key from your secrets manager and then runs backtest360-mcp. A minimal example config is in examples/mcp.json.

Tools

Tool	What it does
engine_info	Engine version, API contract, health
get_me	What the configured key can do: permission scopes, limits, current usage, capability flags
get_catalog	Reference catalogs: operators, execution modes, stop types, sizing methods, bar frequencies, metric sections
list_indicators	Indicator discovery; per-indicator parameter schemas
list_templates	Predesigned strategy templates — discover compactly, fetch one in full, ready to validate and run
get_strategy_schema	JSON Schema for strategy documents
validate_strategy	Validate a strategy without running it — returns structured, locatable errors
run_backtest	Run a historical backtest
get_latest_signal	Evaluate the most recent bar only (no P&L)
compare_backtests	Run several strategies on the same data, side by side
compute_stats	Compute the metric set from an externally produced returns series
search_tickers / list_tickers	Asset discovery for server-side data fetch
get_data_range	Available history and bar-count estimate for a symbol
get_ticker_info	Symbol identity and data coverage in a single call
get_quote	Latest available price for a symbol (paid plan)
get_price_history	OHLCV price history over a date range (paid plan; long histories downsampled to fit)
list_macro_series / get_macro_series	Macroeconomic data: list the series catalog, then fetch one series' observations

The cheap static catalogs are also published as MCP resources (backtest360://catalog/{name}, backtest360://schema/strategy) for clients that support resource attachment.

Prompts

Two workflow prompts scaffold the common multi-tool flows for a connected AI: each names which tools to call, in what order, and what to look at in the results. They carry no interpretation and compute nothing — the connected AI does the reasoning.

Prompt	Arguments	What it scaffolds
robustness_review	symbol, strategy (optional)	Review a backtested strategy for robustness: validate → run → compare against buy-and-hold → weigh the evidence base (sample size, significance/robustness statistics, warnings) → caveated summary
build_and_validate	idea	Turn a plain-language idea into a validated strategy: survey the catalogs → fetch the schema → construct → validate-and-fix loop → dry-run

Response shaping

A full backtest result is megabytes; an agent's context is not. run_backtest and compare_backtests take response_detail:

• summary (default) — headline metrics, warnings, counts, equity endpoints
• stats — every metric the plan allows
• full — plus series (downsampled, endpoints preserved) and trades (paginated)

run_backtest also takes max_series_points (default 500, must be >= 2) to override the series downsampling cap — set it higher for full-resolution series on a long run, or leave it unset for today's default.

include=["trades", "equity_curve", "monthly_returns", "yearly_returns", "signal_diagnostics"] adds specific blocks at any detail level. signal_diagnostics reports which per-bar entry/exit conditions fired, as a capped list of fire dates per condition (not the raw per-bar boolean arrays, which downsampling would corrupt) — or {"available": false, ...} when the run has no condition tree to evaluate (e.g. precomputed signals). Results exceeding the output cap are reduced further and explicitly marked truncated_by_mcp — never silently cut. Shaping only ever selects and thins what the engine returned; no value is computed or altered.

Error semantics

Designed for agents:

• Fixable by changing the request → returned as a normal result: failed validations arrive as {"valid": false, "errors": [...]} with machine codes and document locations; engine rejections arrive as {"accepted": false, "error": ...} with a hint.
• Not fixable that way → a tool error with explicit guidance: rate limits carry the Retry-After value; engine-busy says retry with backoff; a compute timeout says do not retry and reduce scope instead; permission problems name the missing capability. Engine request ids are included for support.

Running the tests (self-host)

pip install -e ".[dev]"
pytest   # unit suite vs a mock engine — no network

Questions / feedback

Questions or feedback? [e-mail] — we read everything. backtest360-mcp is in active development, so help shape it.

Bug reports and feature requests: open an issue on GitHub.

License

MIT — see LICENSE.