echolon
Echolon
Un framework de backtest nativo para agentes LLM destinado a la investigación de futuros. Incluye un servidor MCP, 22 habilidades integradas, 32 códigos de error catalogados y configuraciones Pydantic tipadas; los agentes llaman a herramientas estructuradas en lugar de adivinar las formas de la API a partir de documentación en prosa. De extremo a extremo en futuros diarios de SHFE.
Motor de producción dentro de Qorka, el producto de generación de estrategias nativo de IA de DolphinQuant. Utilizado con dinero real en SHFE cada día de negociación.
Inicio rápido
Tres comandos cubren la trayectoria natural del recién llegado:
Comando | Propósito | Tiempo |
| Demo rápida. Descarga aluminio de SHFE (últimos 2 años) vía akshare, crea el andamiaje de una estrategia y ejecuta un backtest. Requiere red. | ~30s |
| Inicia un proyecto real. Descarga datos de mercado vía akshare (gratis, sin registro), crea el andamiaje de una estrategia a partir de una plantilla y escribe un marcador de espacio de trabajo. | ~1–5 min |
| Itera después de editar. Recorre el árbol para recuperar el contexto del marcador de espacio de trabajo, recalcula indicadores y ejecuta el backtest. No se necesitan flags. | ~5–10s |
pip install echolon
mkdir -p ~/echolon-playground && cd ~/echolon-playground
echolon hello # 30-second demoecholon hello descarga ~2 años de datos de aluminio, crea el andamiaje de la plantilla momentum_breakout, escribe .echolon-workspace.json y ejecuta el backtest. Abre ./echolon-hello/strategy/baseline/entry.py, ajusta un parámetro y luego vuelve a ejecutar con echolon backtest single ./echolon-hello/strategy/baseline/ para ver cómo cambia el Sharpe.
Se incluyen tres plantillas en el paquete: minimal, momentum_breakout, rsi_mean_reversion. echolon examples --list las muestra; pasa --template <nombre> a echolon init / echolon hello para comenzar con una.
Si
pip installfalla en Linux ARM64 / Alpine / FreeBSD, ejecutaecholon doctor: diagnostica la biblioteca C de ta-lib, la única dependencia que puede requerir compilación desde el código fuente fuera de las plataformas estándar de ruedas preconstruidas (Linux x86_64, macOS x86_64+arm64, Windows x86_64; Python 3.11–3.12).
Related MCP server: BotSpot
Contrólalo desde tu agente
pip install echolon # 1. install
claude mcp add -s user echolon -- echolon-mcp # 2. register MCP server (user-wide)
# 3. restart Claude Code to load mcp__echolon__* toolsLuego pregunta:
"Construye una estrategia de seguimiento de tendencia en cobre, haz un backtest de 2018 a 2024."
Detrás de escena, el agente llama a list_skills → elige patterns y quick_start → load_template("momentum_breakout") → list_indicators(has_lookback=True) → edita entry.py y exit.py → ejecuta un bucle validate_strategy_full(strategy_dir) hasta que todo pase → ejecuta el backtest. Si algo falla, analiza [CODE-NNN] del traceback y llama a get_error_doc(code). No hay ningún punto en el que tenga que adivinar.
Entorno de ejecución | Configuración |
Claude Code |
|
Cursor | En |
OpenAI Codex CLI |
|
OpenAI Agents SDK (Python) |
|
LangChain / LangGraph |
|
Cualquier otro cliente compatible con MCP (CrewAI, AutoGen, …) | Configúralo como un servidor stdio con |
Para Claude Code: -s user hace que el registro se aplique a todos tus proyectos (elimínalo para que sea solo para el proyecto actual); -- separa el nombre de registro del comando de lanzamiento. Después de ejecutarlo una vez, claude mcp list debería mostrar echolon como un servidor stdio conectado. La guía de orientación del agente es llms.txt, que también se coloca en la raíz del espacio de trabajo mediante echolon init / hello para que un agente que entre en el proyecto la encuentre sin necesidad del paquete.
Qué está incluido hoy
Hecho de extremo a extremo (grado de producción, ejercitado diariamente):
Investigación de futuros diarios de SHFE: ingesta de datos, catálogo de 214 indicadores, ejecución en Backtrader, optimización TPE de Optuna (objetivo único + multiobjetivo), análisis walk-forward con puntuación de preparación para el despliegue, selección de pruebas robustas basada en KMeans.
Superficie del agente: 23 herramientas MCP, 22 habilidades, 32 códigos de error, 3 plantillas de trabajo.
Aún no (abre un issue si quieres impulsar una parte):
Backtesting intradía de SHFE: pipeline de datos listo, fontanería del motor en proceso de consolidación.
Trading en vivo vía MiniQMT: lanzamiento público limpio en progreso.
Perpetuos de criptomonedas (andamiaje del adaptador CCXT), futuros de CME, acciones.
Alternativas a Optuna (sin grid, sin aleatorio, sin búsqueda de presupuesto bayesiano), orquestación distribuida, Python ≤ 3.10.
Pre-1.0: la API pública puede cambiar entre versiones menores. Los cambios importantes se documentan en CHANGELOG.md.
Trae tus propios datos
Si ya tienes archivos XLS sin procesar de SHFE (descargados de shfe.com.cn), ejecuta SHFEFileDayExtractor directamente en lugar de usar akshare. Para otros formatos (CSV de bróker, tushare, base de datos personalizada), tres archivos deben terminar bajo {workspace}/workspace/data/market_data/SHFE/{instrument}/:
Archivo | Esquema |
|
|
| Mismas columnas, todas las filas concatenadas y ordenadas por fecha. |
|
|
Además, bajo {workspace}/data/SHFE/{instrument_code}/ (ten en cuenta el código CORTO, p. ej., al no aluminum):
Archivo | Esquema |
|
|
Echolon no deriva automáticamente main_contract.csv de OHLCV sin procesar; es una entrada del USUARIO que codifica tu convención de rollover (reglas basadas en volumen, interés abierto o días hasta el vencimiento). Para SHFE vía akshare, echolon init lo deriva por ti; de lo contrario, prodúcelo tú mismo y colócalo en su lugar.
Información del proyecto
Apache 2.0: consulta LICENSE. Úsalo libremente, comercialmente o de otro modo. Desarrollo activo, v0.1.2 beta. Construido y mantenido por DolphinQuant, el mismo equipo que ejecuta Qorka en SHFE. Los issues y pull requests son bienvenidos en github.com/dolphinquant/echolon.
@software{echolon,
title = {Echolon: AI-native quantitative trading engine},
author = {DolphinQuant},
year = {2026},
url = {https://github.com/dolphinquant/echolon},
}Available Tools
23 toolsdescribe_component_apiA
Return the live BaseComponent + IMarketData API surface.
Use this BEFORE writing any component code. The return value is
produced via ``inspect.signature()`` on the actual classes, so it
always reflects the current echolon version — there is no skill /
doc drift. If a method appears here, it exists; if it doesn't, it
doesn't. Pip-install or editable-install, same answer.
Returns a dict with shape::
{
"BaseComponent": {
"override_methods": [{"name": ..., "signature": ..., "doc": ...}, ...],
"helper_methods": [{...}, ...],
"properties": [{"name": ..., "doc": ...}, ...],
},
"IMarketData": {
"methods": [{"name": ..., "signature": ..., "doc": ...}, ...],
},
}
``override_methods`` are the four BaseComponent stubs each component
file must override (``generate_signal`` / ``should_exit`` /
``can_trade`` / ``calculate_size``). ``helper_methods`` are concrete
helpers strategies may call (``get_current_bar``, ``get_indicator``,
``log_*_output``, etc.). ``properties`` are read-only attribute
accessors (``self.market_data``, ``self.portfolio``, ``self.params``).
``signature`` is the formatted ``inspect.Signature`` (e.g.
``"(self, name: str, index: int = 0) -> float"``). ``doc`` is the
first line of the docstring; full docstrings live on the source.
Inputs: no named parameters.
generate_strategy_paramsA
Generate strategy_params.py from params_to_optimize.json.
Deterministic code generation: parses the JSON → determines parameter
ownership across components → emits ComponentParameterTemplate classes
+ framework registration + optuna_search_space with crossover
constraints → writes the target file. Period parameters that exceed
the frequency-appropriate indicator cap are auto-clamped and reported.
Args:
params_file_path: Absolute path to ``params_to_optimize.json``.
output_path: Absolute path to write ``strategy_params.py``.
frequency: ``"interday"`` (caps: TEMA≤62, ADX≤93, default≤180)
or ``"intraday"`` (caps: TEMA≤500, ADX≤750, default≤1000).
Returns ``{"success": bool, "output_path": str, "corrections":
[{"param", "type", "old_*", "new_*", "cap", "category", "changes"?}],
"message": str}``. Parse / IO failures surface as
``success=False`` rather than raising.
Inputs: params_file_path (required), output_path (required), frequency.
get_docA
Read any markdown file under echolon/native/ and return its body.
Generic fallback for content not covered by the dedicated tools
(``get_pattern``, ``get_skill``, ``get_error_doc``, ``load_template``).
Useful for skill-cross-reference resolution and supplementary files
(``echolon/native/skills/SKILLS.md``,
``echolon/native/errors/codes/README.md``, template READMEs).
Args:
path: Relative path under ``echolon/native/`` (e.g.
``"skills/SKILLS.md"``, ``"errors/codes/README.md"``,
``"templates/minimal/README.md"``). Absolute paths and
paths escaping the package root are refused with
``error="path_outside_native"``.
Returns ``{path, body}`` on success, or
``{error, message, path}`` on failure (path-traversal attempt,
file not found, not a markdown file).
Inputs: path (required). Returns: result
get_error_docA
Fetch the structured error documentation for a given error code.
``what`` and ``why`` come from the in-memory registry
(``echolon.errors.ERROR_CATALOG``); the long-form sections (``fix``,
``example``, ``common_causes``, ``related``) come from the per-code
markdown at ``echolon/native/errors/codes/{code}.md``. The full
markdown body is also returned as ``long_form_markdown`` so an
agent can consume the prose verbatim if the parsed sections are
empty (parser-resilience fallback).
Args:
code: Error code like 'VAL-001' or 'IND-003'.
Inputs: code (required).
get_patternB
Return a pattern's structured content, or None if unknown.
Inputs: name (required). Returns: result
get_skillA
Return one skill's body, or None if unknown.
Returns ``{name, description, body, body_no_frontmatter}``.
``body`` is the full file (including YAML frontmatter); use
``body_no_frontmatter`` for the prose alone.
Inputs: name (required). Returns: result
indicator_infoB
Return structured info for one indicator, or None if unknown.
Keys: ``name``, ``has_lookback``, ``function``, ``file``, ``params``
(list of ``{name, default, type}`` dicts). Column names emitted at
runtime by ``processor._build_suffix`` (handles multi-param sweeps).
Inputs: name (required). Returns: result
indicator_paramsA
Return the tunable parameters for an indicator, or None if unknown.
Convenience accessor — same as ``indicator_info(name)["params"]`` but
skips the wrapper dict. Each entry is ``{"name": str, "default": any, "type": str}``.
Inputs: name (required). Returns: result
list_indicatorsA
Return all indicator names in the echolon catalog, optionally filtered by lookback semantics.
Args:
has_lookback: Optional filter.
``True`` → only indicators with a period-like parameter
(e.g. RSI, ATR — sweepable single-dim lookback).
``False`` → only indicators without a period parameter
(no-param indicators like OBV, multi-param scalar indicators
like BBANDS, special-config indicators).
Omit for all.
Inputs: has_lookback. Returns: result
list_patternsA
Return all canonical pattern names.
Inputs: no named parameters. Returns: result
list_skillsA
Return all in-package skills as [{name, description}].
Each skill is one ``SKILL.md`` packet under
``echolon/native/skills/echolon_api/<name>/``. Description comes
from the YAML frontmatter ``description:`` field. Use this as a
directory of "what doctrine is available" before calling
``get_skill(name)`` for the body.
Inputs: no named parameters. Returns: result
list_templatesA
Return all shipped strategy template names.
Inputs: no named parameters. Returns: result
load_templateC
Load a template's files. Returns {'name': ..., 'files': {filename: content}}.
Inputs: name (required). Returns: result
scaffold_componentA
Write a framework-correct scaffold for a strategy component file.
Produces a minimal stub that matches echolon's loader contract —
class name + method signature + return schema — but contains no
trading logic. Coding agents refine the stub into real pathways.
Args:
kind: One of ``"entry"``, ``"exit"``, ``"risk"``, ``"sizer"``, ``"strategy"``.
strategy_dir: Absolute path to the directory where the file is written.
force: If True, overwrite an existing file. If False (default), refuse
and return ``success=False`` with ``error="file_exists"``.
Returns:
{
"success": bool,
"output_path": str (path to the scaffolded file, if success),
"kind": str (echoes the input),
"error": str | None (one of: "unknown_kind", "file_exists", None),
"message": str (human-readable summary),
}
Inputs: kind (required), strategy_dir (required), force.
suggest_similarA
Return up to limit catalog names close to name (difflib + substring).
Inputs: name (required), limit. Returns: result
validate_component_integrationA
Import each component module via StrategyLoader and check its
method arity + strategy_params.DEFAULT_PARAMS top-level shape.
- STR-002: module fails to import (= guaranteed runtime failure).
- PRM-002: DEFAULT_PARAMS missing a required top-level key, or a
value isn't a dict.
- VAL-005: method's required-positional arity (after ``self``)
doesn't match the protocol. Arg NAMES are not checked — the
framework calls positionally, so names are the author's choice.
Returns ``{"any_errors": bool, "findings": [...]}``.
Inputs: strategy_dir (required).
validate_component_loggingB
AST-check that each component's required method calls the
matching self.log_<component>_output(...) with a BaseModel
instance (not a dict, not a wrong schema).
Also flags ``self.params.get(...)`` anywhere in the file (PRM-004 —
defensive dict-access antipattern on the params container).
Returns ``{"any_errors": bool, "findings": [...]}``.
Inputs: strategy_dir (required).
validate_component_protocol_signaturesA
AST-check that each required component class has the required method with a matching return-type annotation (if annotated at all).
- STR-003: class is missing the required method.
- VAL-006: method declares a return annotation but it doesn't match
the expected BaseModel (EntrySignalOutput / ExitSignalOutput / ...).
Missing annotations are NOT flagged (policy vs correctness — missing
annotation is a stylistic choice, Pydantic catches runtime
mismatches). Missing files are silently skipped (preflight STR-001
territory).
Returns ``{"any_errors": bool, "findings": [...]}``.
Inputs: strategy_dir (required).
validate_debug_completionA
Validate post-run debug artifacts landed on disk with the right shape.
Three deterministic checks:
- STR-001: ``selected_robust_trial.json`` + log file both exist.
- VAL-003: artifact JSON parses and carries every required top-level key.
- BT-010: log contains each required marker substring (order irrelevant).
Args:
artifact_path: Absolute path to the JSON artifact (typically
``<workspace>/backtest/selected_robust_trial.json``).
log_path: Absolute path to the debug log file.
required_json_keys: Top-level keys expected on the artifact.
Defaults to ``["trial_number", "params", "metrics"]``.
required_log_markers: Substrings expected in the log. Defaults
to ``["STAGE 4 COMPLETE", "STAGE 5 COMPLETE", "FINAL SUCCESS"]``.
Returns ``{"any_errors": bool, "findings": [{"code", "message",
"context"}, ...]}``. Never raises — parse / file-access failures
surface as findings with the appropriate error code.
Inputs: artifact_path (required), log_path (required), required_json_keys, required_log_markers.
validate_indicator_listA
Validate a flat-dict strategy_indicator_list.json payload against the catalog.
Args:
payload_json: JSON string of the flat-dict
(e.g. ``'{"rsi": {"timeperiod": [10, 20]}}'``).
Returns a dict ``{"valid": bool, "errors": [{code, field, message, suggestion}]}``.
Bad JSON surfaces as a structured error, not a raise.
Inputs: payload_json (required).
validate_parameter_accessA
AST-check for hardcoded threshold literals (PRM-003) and defensive
self.params.get() calls (PRM-004) across the 4 component files.
Highest FP-risk of the B1 validators — PRM-003 findings include
``context["severity"] = "warning"`` so callers can treat them as
non-blocking while the allowlist is tuned. PRM-004 findings are
always bugs (framework contract violation).
Allowlist covers: range() args, index slicing, keyword args, default
args, None comparisons, small integer constants ``{-1, 0, 1, -1.0, 0.0, 1.0}``,
and string-literal comparisons (framework-defined regime / signal enums).
Returns ``{"any_errors": bool, "findings": [...]}``.
Inputs: strategy_dir (required).
validate_strategyB
Validate a strategy directory against echolon contracts.
Args:
path: Absolute path to the strategy directory.
Returns a dict with keys: status ("VALID"|"INVALID"), errors (list of dicts).
Inputs: path (required).
validate_strategy_fullA
Run every shipped validator and return merged findings.
Composes the individual MCP validators (``validate_strategy`` /
``validate_component_protocol_signatures`` /
``validate_component_integration`` / ``validate_component_logging``
/ ``validate_parameter_access``) so an agent gets a complete
validation report from a single tool call.
Args:
strategy_dir: Absolute path to the strategy directory.
Returns:
``{status, any_errors, total_findings, findings: [{code, ...}],
invocations: [{validator, count}]}``. ``status`` is ``"VALID"``
iff no findings were reported by any validator.
Inputs: strategy_dir (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 23 tools
Each tool has a clearly distinct purpose: API introspection, parameter generation, documentation retrieval, catalog listing, scaffolding, and various validation checks. Overlap among validation tools is resolved by specific validation targets (integration, logging, protocol, parameters, etc.).
Tools follow a mostly consistent verb_noun pattern (e.g., get_doc, list_indicators, validate_strategy). Minor deviations like 'indicator_info' instead of 'get_indicator_info' and 'suggest_similar' instead of 'get_suggestions' are tolerable.
23 tools cover a comprehensive set of operations for a strategy framework: API description, parameter generation, documentation, catalog browsing, scaffolding, and multiple validation checks. The count is well-scoped for the domain.
The tool surface covers key development and validation workflows. Minor gaps exist (e.g., no tool for creating a new strategy directory or running backtests), but the core lifecycle of building and validating components is well-supported.
Maintenance
Related MCP Connectors
Trading MCP server for AI agents, with live market data, account reads and controlled execution.
MCP server exposing the Backtest360 engine API as tools for AI agents.
MCP server for OpenMM — exposes market data, account, trading, and strategy tools to AI agents
Research-only MCP server: your AI as a quant research desk. 90 tools, no trades, no brokers.
Related MCP Servers
- AlicenseBqualityDmaintenanceAn extensible framework that exposes quantitative research functions and financial data connectors, such as FRED, via an MCP server. It enables users to perform complex financial modelling, data retrieval, and autonomous research loops with built-in guardrails and pluggable components.64MIT
- FlicenseNot gradedqualityBmaintenanceFull-lifecycle algorithmic trading MCP server. AI strategy generation from plain English, backtesting, live bot deployment to 10+ brokers, portfolio monitoring, and prediction markets. Stocks, options, crypto, futures. 32 tools. Free tier.-
- AlicenseNot gradedqualityDmaintenanceLocal-first backtesting engine with built-in overfitting detection (PBO, deflated Sharpe, bootstrap CI, walk-forward) and a native MCP server for AI agents to validate trading strategies.4Apache 2.0
- AlicenseAqualityBmaintenanceMCP server for the FLOX trading framework. About 30 tools to run backtests, scaffold strategies, validate for lookahead bias, compute indicators, place orders, and query PnL from Claude/Cursor.38945 PyPI231MIT