FLOX C++ trading systems framework with MCP

v0.6.0 — polyglot strategies, AI-friendly DX, same code backtest→live

Strategies in Python, Node.js, Codon, and embedded JavaScript run on one C++ core through parity-tested bindings. The same strategy code goes from backtest to live without a rewrite — pre-trade gate, fill hooks, order events all behave the same on both runners. flox-mcp is a Model Context Protocol server that lets coding agents query the same FLOX install you're using: every binding's surface, every indicator, every error code, the live engine's positions and orders, the full doc set.

Highlights

• flox-mcp MCP server. Install with pip install flox-mcp, bootstrap with flox-mcp init. Tools cover symbol lookup across bindings, indicator introspection, scaffolders, full-text doc search, runtime inspection, control-plane order routing, live analytics.
• Same code backtest→live. BacktestRunner now takes the same RiskManager / KillSwitch / OrderValidator / PnLTracker stack as the live Runner. Reduce-only and flatten orders bypass the gate (so you can always exit when caps tighten).
• flox engine sim paper engine in one command. Boots Runner + SimulatedExecutor + ControlServer + a runtime-snapshot writer. Prints the engine URL and a copy-pasteable flox-mcp init --engine-url ... --token ....
• Polyglot parity, all four bindings. Multi-TF alignment helpers, composite-condition DSL, indicator-grid sugar, OrderGroup state machine, MultiFeedClock, .floxrun trace format, latency models, tape diff, portfolio risk, execution algorithms. Cross-binding parity gate in CI.
• Strategy.on_fill / on_order_update hooks across all bindings. Strategies see their own fills and order-lifecycle events.
• .floxlog public spec + replay-equivalence CI gate. The recorded tape format is documented; a CI step blocks any binding change that breaks byte-for-byte replay equivalence.
• AI-DX surface. llms.txt and llms-full.txt published, .pyi + py.typed on PyPI, index.d.ts on npm, structured FloxError with code + help URL, snippet linter that fails when doc examples drift.
• IDL-driven C API codegen. flox_capi_spec.hpp is what generates the C header, the Codon FFI, and the LLMs reference. ABI snapshot gate fails the build on accidental breakage.

What's new

AI / MCP

• flox-mcp Model Context Protocol server (#141)
• Polyglot tools: lookup_symbol, list_bindings over bundled cross-binding data (#166)
• Runtime tools: run_backtest, compute_indicator, suggest_indicator (#168)
• Read-only positions inspection (#174)
• Mutating control-plane tools with capability scopes + token security model (#178)
• Live analytics: read-only introspection over the control plane (#179)
• Lookahead-bias detector tool (#181)
• lookup_symbol surfaces curated semantic gotchas (#229)
• record_data tool wraps the canonical recording paths (#228)
• init_project thin wrapper around flox new (#225)
• flox-mcp init — bootstrap .mcp.json in one command (#233)
• Tier-4 read tools fail-graceful when no runtime snapshot exists (#232)
• docs_search AND-matches multi-word queries (#221)
• run_backtest routes by strategy kind (#222)
• suggest_indicator covers mean-reversion / stddev / z-score (#223)
• Friendlier no-engine error for runtime tools (#224)
• Surface canonical paths instead of building parallel ones (#219)

Strategy & backtest

• BacktestRunner pre-trade gate parity with live Runner (#235)
• Strategy.on_fill / on_order_update hooks — engine + pybind11 (#226)
• BacktestRunner.run_tape(path) for .floxlog replay (#227)
• Wire MultiModePositionTracker into Strategy (silent zero-trade fix) (#220)
• Strategy hot-reload without dropping connections (#192)
• Multi-TF alignment helpers on Strategy (#196)
• Strategy.on_bar callback + run_bars (#123)
• RL environment over a .floxlog tape, Gymnasium-compatible (#186)
• Execution algorithms: TWAP / VWAP / Iceberg / POV slicers (#185)
• PaperBroker: live-data sandbox / paper trading mode (#175)

Polyglot parity

• Cross-binding parity gate in CI (#149)
• TypeScript strategy authoring with decorators (#187)
• Latency models, tape diff, portfolio risk, execution algos — C++ core with binding parity (#188, #189, #190, #191)
• .floxrun strategy-trace format with cross-binding parity (#194)
• Composite-condition DSL for multi-symbol multi-TF strategies (#197)
• Indicator-grid sugar on top of composite DSL (#198)
• Latency-aware MultiFeedClock (#200)
• Multi-leg OrderGroup helper, BestEffort policy (#199)
• Multi-TF alignment helpers in QuickJS + Codon (#204)
• Composite DSL ported to NAPI / QuickJS / Codon (#205)
• Indicator-grid sugar ported to NAPI / QuickJS / Codon (#206)
• OrderGroup state machine across all four bindings (#207)
• MultiFeedClock state machine across all four bindings (#208)
• Runner cross-binding .floxrun parity gate (#212)
• Strategy.on_fill / on_order_update parity for NAPI / QuickJS / Codon (#230)
• BacktestRunner.runTape — NAPI / Codon parity (#231)

Tooling & CLIs

• flox engine sim — paper engine in one command (#234)
• flox tape CLI: record / replay / inspect (#173)
• flox bundle: reproducibility bundle format (#176)
• flox tape diff: trade-by-trade comparison of two tapes (#182)
• flox lint lookahead (#181)
• flox new CLI scaffolder (#158)
• CcxtBroker live broker adapter (#157)
• Notebook variant of the research scaffold (#160)
• Live-trading template via flox new --template=live (#162)
• Indicator-library scaffolder (#165)

Indicators & analytics

• Forward-looking targets distinct from indicators (#115)
• ADF stationarity test (#116)
• AutoCorrelation indicator (#117)
• IndicatorGraph (batch) (#118)
• Streaming step()-based IndicatorGraph for live trading (#119)
• Single source of truth for indicators and DAG (#122)
• HTML backtest report + equity_curve / trades accessors (#161)
• Walk-forward + type-erased grid search across bindings (#163)
• Heatmap HTML renderer across all bindings (#164)
• MLflow integration for FLOX backtests (#167)
• White's reality check across all bindings (#169)
• Latency models for backtest realism (#184)
• MVP web replay viewer for tape and run artifacts (#195)

Replay / tape format

• Public specification of the .floxlog binary tape format (#180)
• Replay-equivalence CI gate (#177)
• Delta-book compression in the tape format (#193)
• Live-tail safety + book-update reads (#120)
• Mid-stream _from seek across all bindings (#126)
• LZ4 link in the Node binding so compressed segments read on every platform (#127)
• TraceSignalHandler / TraceExecutionListener adapters for .floxrun (#201)
• flox bundle ships expected.floxrun alongside expected_output.json (#202)
• Runner attaches a TraceRecorder for one-call signal auto-capture (#209)
• Trace order events and fills through the attached recorder (#210)
• OrderGroup auto-dispatches recommended actions through the executor (#211)
• OrderGroup basket-level risk gate before submission (#215)
• OneSided pair-latency budget on leg-A ack (#216)
• Bar-close dispatch ordering cross-binding parity test (#217)

Errors, typing, docs

• llms.txt and llms-full.txt published for AI agents (#128)
• .pyi stubs + py.typed for typed PyPI distribution (#129)
• index.d.ts for typed npm distribution (#130)
• Structured FloxError with code + help URL (#138)
• Auto-generated Python API index from .pyi (#139)
• Snippet linter + ratchet allowlist (#140)
• Top-20 throw sites migrated to FloxError (#154)
• Cross-binding FloxError surface in Node (#155)
• Snippet coverage ratchet enforcement (#156)
• First-backtest tutorial with real BTC sample (#159)
• Architecture + contributor guide (#152)
• Cross-symbol pairing for two-input indicators (Correlation, AutoCorrelation) (#237)
• mcp/README.md flox-mcp init quickstart (#238)
• README hero — polyglot + AI-friendly + same-code backtest-to-live (#240)
• Python + Node package READMEs refreshed (#241)
• pre-commit hook auto-regens .pyi / llms.txt / mcp data (#239)

C-API / IDL codegen

• IDL-driven C API codegen — slice prototype (#131)
• Full-coverage spec + enums/aliases + bootstrapper (#132)
• emit-codon + emit-llms emitters (#133)
• ABI snapshot CI gate (#135)
• RiskManager pre-trade hook (#136)
• KillSwitch + OrderValidator pre-trade hooks (#137)
• Logger callback hook for binding-side log redirection (#144)
• Post-emission observer hooks: PnLTracker, StorageSink (#145)
• MarketDataRecorder hook (#146)
• ReplaySource hook for binding-supplied event streams (#147)
• Executor + ExecutionListener hooks (#148)
• pybind11 wrappers for extension hooks (#150)
• NAPI wrappers for extension hooks (#151)
• Pointer-variant create exports unblocking pure-Codon strategies (#213)

Connectors / build

• Native connectors merged into flox under FLOX_BUILD_CONNECTORS (#171)
• CMake feature flags: FLOX_ENABLE_* → FLOX_BUILD_* split + CI flag matrix (#172)

Full changelog: https://github.com/FLOX-Foundation/flox/compare/v0.5.6...v0.6.0