Quant Framework

An open, pluggable framework for composable quantitative workflows. Start with FRED. Expand to anything.

Inspired by Karpathy's autoresearch — the same three-layer contract (immutable evaluator, agent sandbox, human direction), applied to quantitative finance as an extensible framework.

This is a framework — not a product. FRED is the hello-world connector. Everything else is an extension of the same pattern.

Prerequisites

• Python 3.12.10+
• uv — Python package manager
• FRED API Key — Get one free from FRED

Installation

# Clone the repository
git clone <repo-url>
cd quant_framework

# Install all dependencies
uv sync

Configuration

Environment Variables

Create a .env file in the project root (or export directly):

# .env
FRED_API_KEY=your_api_key_here

Persona Config

Edit configs/persona.yaml to control which functions and connectors your MCP server exposes:

name: "Quant Research Agent"
description: "MCP server exposing quantitative research functions"
host: "127.0.0.1"
port: 8000

functions:
  - run_linear
  - run_random_forest
  - run_svr
  - run_xgboost
  - run_bayesian_ridge
  - run_hmm

connectors:
  - fred

Guardrails Config

Edit configs/guardrails.yaml to define validation rules for function outputs:

defaults:
  max_records: 10000

rules:
  run_linear:
    max_records: 5000
    required_fields: [model, r_squared, coefficients]
    roles:
      analyst:
        redacted_fields: [model]

Usage

CLI — Start the MCP Server

# Show available commands
uv run quant --help

# Start the MCP server with SSE transport
uv run quant serve --persona configs/persona.yaml

# Use stdio transport instead
uv run quant serve --persona configs/persona.yaml --transport stdio

This will:

1. Register all modelling functions from the FunctionRegistry
2. Initialise connectors (auto-connects using $FRED_API_KEY)
3. Start the MCP server on 127.0.0.1:8000

Connect from Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "quant-framework": {
      "url": "http://localhost:8000/sse"
    }
  }
}

Run the Example Script

uv run python examples/basic_usage.py

This demonstrates:

1. Querying GDP data from FRED
2. Running linear regression via the FunctionRegistry
3. Validating the result through the GuardrailEngine

Project Structure

quant_framework/
├── pyproject.toml                # Dependencies & CLI entry point
├── configs/
│   ├── persona.yaml              # MCP server persona config
│   └── guardrails.yaml           # Validation rules
├── examples/
│   └── basic_usage.py            # End-to-end demo script
├── experiments/                  # Autonomous research loop files
│   ├── evaluate.py               # Evaluation harness (scalar metric)
│   ├── prepare_snapshot.py       # Data snapshot caching script
│   └── strategy.py               # Editable strategy sandbox
├── program.md                    # Human-directed research agenda
└── quant_framework/              # Package root
    ├── cli.py                    # CLI (quant serve)
    ├── core/
    │   ├── function.py           # @register_function, FunctionRegistry, FunctionResult
    │   └── guardrail.py          # GuardrailEngine, GuardrailViolation
    ├── connectors/
    │   ├── connectors.py         # BaseConnector, ConnectorRegistry
    │   └── fred.py               # FREDConnector (with 24h file cache)
    ├── functions/
    │   └── modelling.py          # Registered modelling functions
    └── mcp/
        └── generator.py          # MCPServerGenerator

Core Components

Connectors

from quant_framework.connectors import FREDConnector

fred = FREDConnector()
fred.connect({"api_key": "your_key"})
df = fred.query("GDP", observation_start="2020-01-01")

Modelling Functions

All functions are registered with @register_function and return a FunctionResult:

from quant_framework.functions.modelling import run_linear_regression

result = run_linear_regression(df, target="GDP", features=["UNRATE", "FEDFUNDS"])
print(result.output["r_squared"])   # 0.12
print(result.trace_id)              # unique trace ID

Guardrail Engine

from quant_framework.core import GuardrailEngine

engine = GuardrailEngine("configs/guardrails.yaml")
engine.validate("run_linear", result.output)           # passes
engine.validate("run_linear", result.output, role="analyst")  # applies role-specific rules

• Hot-reload: edits to the YAML take effect immediately (checks file mtime)
• Per-role overrides: stricter rules for specific roles

Function Registry

from quant_framework.core import FunctionRegistry

# List all registered functions
FunctionRegistry.list()                        # ['run_linear', 'run_random_forest', ...]
FunctionRegistry.list_by_category("modelling") # filter by category

# Call by name
result = FunctionRegistry.call("run_linear", df=df, target="GDP")

The Autonomous Research Loop

The framework includes a fully autonomous research loop designed to test hypotheses and incrementally improve a quantitative strategy.

It builds on the three-layer contract outlined in program.md:

1. Fixed Evaluation Harness (experiments/evaluate.py): Scores the strategy on a fixed historical dataset.
2. Strategy Sandbox (experiments/strategy.py): The single file where the agent tests features, model choices, and signal logic.
3. Human Direction (program.md): Defines the agent's constraints and the high-level research agenda.

Running the Loop

Provide the program.md file to any autonomous coding agent (like Claude or the built-in system) and instruct it to begin. The agent will read program.md, modify experiments/strategy.py, run evaluate.py, and use a keep/discard ratchet to only commit changes that improve the composite score.

Extending the Framework

Add a Connector

from quant_framework.connectors.connectors import BaseConnector, ConnectorRegistry

@ConnectorRegistry.register("bloomberg")
class BloombergConnector(BaseConnector):
    def connect(self, config): ...
    def query(self, request, **kwargs): ...
    def get_schema(self): ...
    def health_check(self): ...

Add a Function

from quant_framework.core import register_function, FunctionResult

@register_function(name="my_indicator", category="technical")
def my_indicator(df, window=14):
    result = ...  # your logic
    return FunctionResult(output={"value": result}, metrics={"window": window})

The function is automatically available in the FunctionRegistry and can be exposed as an MCP tool by adding its name to your persona YAML.

Design Principles

• Connector-first. Every data source is a BaseConnector. Learn one interface, connect anything.
• Functions as atoms. Decorated Python functions that auto-register and auto-expose via MCP.
• Progressive complexity. Start with FRED. Add what you need, when you need it.
• Three-layer contract. Immutable evaluator (guardrails), agent sandbox (function store), human direction (persona configs).

Contributors

Arjun Singh

License

MIT